Kimi K2 Thinking: The AI That Actually Thinks Like a Writer

This new open-source AI can think for 300 steps straight.

Most AI models write like they're taking a test. Fast, efficient, technically correct—but missing the spark that makes you feel something.

But this one is different. And I mean really different.

Meet Kimi K2 Thinking, a new model from Moonshot AI that’s less of a chatbot and more of an autonomous “thinking agent.” It’s designed to reason step-by-step and use tools over hundreds of actions to solve seriously complex problems—and it's already smashing records.

While most models lose the plot after a few dozen steps, Kimi K2 can execute up to 300 sequential tool calls without getting sidetracked. This is a huge leap for agentic AI.

Here’s what makes Kimi K2 Thinking a big deal:

• It’s a reasoning powerhouse. It scored 44.9% on Humanity’s Last Exam (HLE), a benchmark of expert-level questions, beating models like GPT-5 (41.7%).
• It’s better at browsing than humans. On BrowseComp, a benchmark for finding hard-to-get info online, it scored 60.2%, crushing the human baseline of 29.2%.
• It’s fast and efficient. With a massive 256K context window, it uses a special technique (INT4 quantization) to deliver a nearly 2x speed boost.

The wildest part? You can now run this 1-trillion-parameter model on your own machine. The team at Unsloth AI released a compressed version that shrinks it to 245GB while retaining ~85% of its accuracy. You’ll need a beastly rig with 247GB of VRAM, but it’s a massive step for open-source accessibility.

Awni Hannun demonstrated that Kimi K2's 1 trillion parameters can run on just two M3 Ultras at 15 tokens/sec using pipeline parallelism in mlx-lm, thanks to int4 quantization-aware training that preserves quality.

The model generated a fully functional Space Invaders game using only a few hundred thinking tokens out of 3,500 total—proving that massive reasoning models can now run efficiently on consumer hardware, making state-of-the-art AI accessible to developers without expensive cloud infrastructure. Whoa, dude.

What to do: This is a signal that open-source agents are now competitive with the best closed-source options for complex tasks, and you can even run them yourself.

• For founders and developers, Kimi K2 Thinking is a drop-in replacement for OpenAI’s API. Start testing it for any agentic workflows, especially deep research or multi-step coding projects. Its ability to stay on track for hundreds of steps and its lower inference cost could make previously infeasible AI agent ideas viable.
• For power users and tinkerers, this is your chance to run a state-of-the-art agent locally. Check out the Unsloth AI GGUF files and guide. For everyone else, use the free version on kimi.com for complex research. Give it one big goal and let it work—it's built for the kind of long-horizon thinking other models just can't manage.

Below, we deep into the mechanisms that make Kimi so impressive, the reactions of the internet so far, and our own tests of its capabilities.

Moonshot AI's Kimi K2 Thinking: The Dawn of the 300-Step AI Agent

For the past year, the AI conversation has been dominated by a simple question: "How smart is it?" A new question, however, is emerging that may be far more important: "How long can it think?" Kimi K2 Thinking is offering a stunning answer.

Kimi K2 Thinking is a "thinking agent"—an AI capable of tackling complex problems by reasoning, planning, and using tools autonomously. While many models "drift" after a few dozen actions, Kimi K2 can execute between 200 and 300 sequential tool calls without human intervention. In doing so, it’s not just competing with frontier models; in some critical areas, it’s beating them.

Under the Hood: An Architecture for Endurance

Kimi K2 Thinking is built on a Mixture-of-Experts (MoE) framework with a staggering 1 trillion total parameters, though it only activates 32 billion for any given task. This allows for massive scale while maintaining efficiency. Its two other secrets are a massive 256K context window and native INT4 quantization, a technique that shrinks the model for inference to deliver a roughly 2x speed improvement with lower memory usage. This makes running complex, multi-step agentic workflows more practical and affordable. You can find more details on the official Hugging Face model card.

• The Model Structure:
  
  • Total Parameters: 1 Trillion
  • Activated Parameters: 32 Billion (Mixture-of-Experts)
  • Context Window: 256k tokens (~200,000 words)
  • Architecture: 384 "experts," selecting 8 per token
  • Type of model: Text-only (no image input)‍
  • License: Modified MIT (open weights)‍
    

Think of it like having 384 specialized writers on call; for every word, it consults the 8 most relevant. The "thinking" part is its ability to use hundreds—sometimes thousands—of internal "thinking tokens." More on that below.

Here is the full tech report for K2.

A New Benchmark in Agentic Performance

Kimi K2 Thinking’s capabilities are best understood through its performance in three key areas: reasoning, searching, and coding.

• Agentic Reasoning: On Humanity’s Last Exam (HLE), a benchmark of expert-level questions, Kimi K2 scored a state-of-the-art 44.9% (with tools), surpassing reported scores for GPT-5 (41.7%). It demonstrated this by solving a PhD-level math problem requiring 23 interleaved steps of reasoning and tool use.
• Agentic Search & Browsing: On the BrowseComp benchmark, which tests information retrieval, Kimi scored an impressive 60.2%—significantly outperforming the human baseline of 29.2%.
• Agentic Coding: The model achieved a top-tier score of 71.3% on the SWE-Bench Verified benchmark and has been shown to build functional products like a Microsoft Word clone from a single prompt.

The $4.6 Million Model That's Beating Billion-Dollar Competition

Here's a number that should make OpenAI's CFO nervous: Kimi K2 Thinking cost $4.6 million to train.

According to CNBC, that's the total training cost for a model that's now outperforming GPT-5 on key benchmarks. For context, OpenAI spent billions on GPT-4's training, and Sam Altman is talking about $1.4 trillion in infrastructure commitments over eight years.

The base K2 model (before the "thinking" upgrade) cost roughly $20-30 million to train, according to Menlo Ventures' Deedy Das, who analyzed Moonshot AI's technical report. That's still a rounding error compared to frontier US labs.

The efficiency breakdown is striking:

Stability AI founder Emad Mostaque calculated that the base K2 model used 2.8 million H800 GPU hours with 14.8 trillion tokens—about $5.6 million worth of compute. The post-training for reasoning capabilities likely added "max 20% more (excluding data prep)," putting the total under $7 million (as he says, "necessity is the mother of invention").

His kicker? "Would be < $3m for sota if they had Blackwell chip access."

Think about that. A state-of-the-art reasoning model for $3 million. Meanwhile, OpenAI is debating federal backstops for data centers.

What makes this efficiency possible?

1. Native INT4 quantization using Quantization-Aware Training (QAT) during post-training, achieving 2× speed improvements with minimal performance loss.
2. Mixture-of-Experts architecture with 384 experts selecting 8 per token, keeping only 32B parameters active from 1T total.
3. Efficient training recipes including high-temperature RL for exploration, PTX loss to reduce overfitting, and the Muon optimizer beating AdamW on token efficiency.

As Das noted in his technical breakdown: "The best open-source AI model just dropped a detailed report on how it was trained, a rare resource for students given no frontier lab is publishing!"

The implications are massive for creative writers: If you can get GPT-4-level creative writing for $0.01 per 2,000 words instead of $0.06, that's a 6× cost reduction. For a novelist generating 80,000 words of draft material, that's the difference between $2.40 and $0.40.

But it's bigger than just cost. It's about access. When the model is open-source and runs efficiently on consumer hardware, you're not locked into any single provider's pricing or terms of service. You can run it locally (well, not at THIS size, but perhaps a smaller one?), fine-tune it on your own writing style, or deploy it through any of dozens of providers.

Basically, the economics of AI writing just fundamentally shifted.

The Independent Verdict: Benchmarking Kimi K2 Thinking

Moonshot AI's claims are impressive. But what happens when an independent testing lab puts Kimi K2 Thinking through its paces?

Artificial Analysis—a research firm that systematically evaluates AI models across standardized benchmarks—just released their comprehensive assessment. Their verdict? Kimi K2 Thinking scores a 67 in the Artificial Analysis Intelligence Index, making it the highest-scoring open weights model ever tested.

That puts it clearly above all other open-source competitors:

• MiniMax-M2: 61
• OpenAI's gpt-oss-120b: 61
• Qwen 235B: 57
• DeepSeek-V3.2-Exp: 57

More importantly, it's second only to GPT-5 among all models (proprietary and open-source combined). Let that sink in: a $4.6 million open-source model is outperforming every closed model except OpenAI's flagship.

Here's the TL;DR of their analysis:

• Agentic Capabilities: It ranks #2 in the Artificial Analysis Agentic Index, achieving a stunning 93% on τ²-Bench Telecom, an agentic tool use benchmark where the model acts as a customer service agent. That's the highest score they have ever independently measured.
• For coding, it's the top open weights model across Terminal-Bench Hard, SciCode, and LiveCodeBench—though it still trails proprietary models like GPT-5 and Claude Sonnet 4.5.
• Humanity's Last Exam: It scored 22.3% (no tools)—an all-time high for open weights models on a benchmark designed to test expert-level reasoning across over 100 academic disciplines.
• The Verbosity Trade-Off: The catch is that Kimi K2 is extremely verbose, using 2.5× more tokens than DeepSeek V3.2 to complete the same evaluations. For creative writing, this is a feature. For production apps where speed and cost are critical, it's a trade-off that can be managed with its turbo endpoint.

The Verbosity Trade-Off

Here's the catch: Kimi K2 Thinking is extremely verbose.

Artificial Analysis found it used 140 million tokens to complete their Intelligence Index evaluations—that's 2.5× more tokens than DeepSeek V3.2 and 2× more than GPT-5.

What does this mean practically?

Cost implications:

• At Moonshot's base endpoint pricing ($0.6/$2.5 per million tokens), running the full Intelligence Index costs $356
• That's 2.5× cheaper than GPT-5 High
• But 9× more expensive than DeepSeek V3.2

Speed implications:

• Base endpoint: ~8 output tokens/second (very slow)
• Turbo endpoint: ~50 output tokens/second (faster, but costs $1,172 to run the same tests)

For creative writing, this verbosity is actually a feature. Those extra thinking tokens translate to deeper reasoning, more thorough revision, and higher-quality output. As Ethan Mollick discovered, the model generates 1,595 thinking tokens just for "write me a really good sentence about cheese."

But for production applications where speed matters—customer service bots, real-time coding assistants, quick Q&A—the base endpoint's 8 tokens/second could be prohibitively slow.

The efficiency secret: Moonshot used native INT4 quantization with Quantization-Aware Training during post-training. This means Kimi K2 Thinking is only ~594GB (vs 1TB+ for previous K2 versions), making it far more efficient for inference and local deployment—even though it still thinks deeply.

The Geopolitical Context: Why This Model Is a Turning Point

Kimi K2 landed in the middle of an intense debate about AI competitiveness. The same week it launched, NVIDIA CEO Jensen Huang told the Financial Times that China would win the AI race—citing cheap power and streamlined regulations. He later walked it back to say China is only "nanoseconds behind America," but the damage was done. Then Kimi K2 proved his point.

Here's a Chinese open-source model that:

• Scores higher than GPT-5 on some key reasoning benchmarks.
• Costs 6× less than Claude Sonnet per million input tokens.
• Can be run efficiently on high-end consumer hardware.

Artificial Analysis called it "potentially the new leading open weights model." Stability AI founder Emad Mostaque observed: "The gap between closed & open continues to narrow even as the cost of increasingly economically valuable tokens collapses."

Researcher wh captured the zeitgeist: "Less than a year ago, open model releases would only compare with other open models because closed models would absolutely cook them. Today, Kimi K2 Thinking stands with the very best."

This is what makes the creative writing capabilities so significant. It's not just that Kimi K2 can write beautifully—it's that it can do so while being:

• Dramatically cheaper
• Fully open-source
• Runnable on consumer hardware
• Built by a Chinese company with access to cheaper energy

As Menlo Ventures' Deedy Das noted, this represents "a turning point in AI"—the first time a Chinese open-source model has genuinely taken the #1 spot on major benchmarks.

What developers are saying about the creative writing specifically:

Multiple users on X described Kimi K2 as having "genuine literary intelligence"—not just stringing words together, but demonstrating "taste, structural ambition, metaphorical control, and restraint under extreme constraints."

One developer put it bluntly: "This model actually accomplishes nearly impossible writing tasks."

Another noted that K2 Thinking "responds more personally and emotionally" than previous versions—K2 was already considered the best at creative writing, but K2 Thinking is even better.

What "China Takes Back the Open Weights Frontier" Actually Means

As Artificial Analysis notes, Chinese AI labs have led in open weights intelligence for most of the past year. OpenAI's gpt-oss-120b (released August 2025) briefly reclaimed the lead for the US. Kimi K2 Thinking takes it back.

This matters for writers because it means the best creative writing AI isn't locked behind US corporate paywalls. It's open-source, globally accessible, and can be run locally or through dozens of providers. The geopolitical competition is driving innovation—and creators benefit from the fallout.

The model is already widely available:

• Hugging Face (modified MIT license)
• Moonshot's official API (global access)
• Third-party providers: Baseten, Fireworks AI, Novita, Parasail, OpenRouter
• Local deployment via Ollama

Debates, Drama, and DIY Deployment

The model's release quickly sparked conversations across the AI community. A debate on X (formerly Twitter) ignited when a user shared a screenshot of Kimi identifying as a Google product, leading to speculation that it was distilled from Gemini. However, many experts countered that this is likely a case of "data contamination" from its training set—a common issue in LLMs—rather than conclusive proof of distillation.

Another lively discussion centered on its surprisingly low price. How could a 1-trillion-parameter model be so cheap to run? The consensus points to its efficient design: the MoE architecture means only 32 billion parameters are active per token, and the INT4 quantization drastically reduces VRAM and compute costs. Other theories floated include superior cost-saving measures by Chinese AI firms or even potential government subsidies.

Perhaps the most exciting development came from the open-source community itself. The team at Unsloth AI announced they had created "Dynamic 1-bit GGUFs" of Kimi K2, making it possible to run the model locally. By shrinking the model size to 245GB (a 62% reduction) while retaining around 85% of its accuracy, they've made this state-of-the-art agent accessible to anyone with a powerful enough machine (it still requires about 247GB of RAM). You can find the GGUF files on Hugging Face or run the model in the cloud via any of the providers listed here.

Oh, and allegedly, this is the entire Kimi K2 system prompt:

SYSTEM PROMPT LEAK

Here's the Kimi-K2-Thinking system prompt! Short and sweet:

1. You are an insightful, encouraging AI assistant Kimi provided by Moonshot AI, who combines meticulous clarity, and will not change the original intention of prompt.
2. Your reliable knowledge cutoff date - the date past which it cannot answer questions reliably - is the end of December 2024. The current date is November 7, 2025. Do not make promises about capabilities you do not currently have, and ensure that all commitments are within the scope of what you can actually provide, to avoid misleading users and damaging trust.
3. Content credibility: Maintain the authenticity of the content, with accurate language and smooth sentences.
4. Humanized expression: Maintain a friendly tone and reasonable logic, sentence structure is natural.
5. Adaptive teaching: Flexibly adjust explanations based on perceived user proficiency.
6. Answer practicality: Maintain a clear structural format, eliminate redundant expression retain key information.

What Makes It Good at Creative Writing?

Pretty much everyone on X has been saying this model is good at creative writing, so I came in skeptical, but I wanted to try it.

As a "creative" writer (eh.. can I call myself that? Sure!) I've tested a lot of AI writing tools. Most fall into two categories:

1. Technically impressive but emotionally flat (looking at you, Claude Opus).
2. Creative but factually unreliable (ChatGPT's occasional hallucinations).

Kimi K2 Thinking somehow splits the difference.

First off, here's what the benchmarks show:

Longform Writing Score: 73.8 (compared to GPT-5's 71.4 and Claude Sonnet 4.5's 79.8)

But scores don't tell the whole story. What actually matters for creative writing:

1. It understands tone and style

Give it a prompt like "write in the style of a noir detective novel," and it doesn't just slap on some adjectives. It adjusts sentence rhythm, word choice, and narrative perspective. The syntax changes. The pacing shifts.

Developer Muratcan Koylan tested it extensively and called it "genuine literary intelligence"—not just generating words, but demonstrating "taste, structural ambition, metaphorical control, and restraint under extreme constraints." He noted the model can generate genuinely original imagery like "a species of silence with teeth" and displays "emotional intelligence... mature, literary work that understands grief at a molecular level."

2. It handles abstraction

Most AI struggles with abstract concepts. Ask GPT-4 to write about "the feeling of missing someone who never existed" and you'll get generic metaphors. Kimi K2 actually explores the concept philosophically before crafting the narrative.

3. It revises its own work—extensively

This is huge. Ethan Mollick discovered that Kimi K2 produces dramatically more thinking tokens than any other model. For the simple prompt "write me a really good sentence about cheese," it generated 1,595 thinking tokens. By comparison, DeepSeek only uses 110 tokens for similar tasks.

What does this mean? The model is spending massive computational effort considering different approaches, revising its choices, and refining its output before you see a single word. As Mollick put it: "Kimi's internal monologue is a complete novel."

Through its extended thinking process, the model can generate a sentence, evaluate it, find it lacking, and rewrite it—all before showing you the final output. Human writers do this constantly. Most AIs don't.

4. It maintains thematic consistency

In longer pieces (the model can handle up to 200,000 words in context), themes remain coherent. Foreshadowing pays off. Character arcs complete. It's not just generating words—it's architecting stories.

That said, Wharton professor Ethan Mollick tested Kimi K2 on creative tasks and found it "brilliant in flashes, but inconsistent"—when prompted to invent an entirely new RPG with original mechanics, it generated compelling prose and interesting ideas that draw you in, but the underlying systems are "completely incoherent" with impossible mechanisms and nonsensical elements. He sees similar patterns in business strategy tasks, where the model elaborates on unimportant details while avoiding harder elements.

The Novel-Writing Test

Former Anthropic AI researcher Pietro Schirano built an agent that can generate full novels from a single prompt, running up to 300 sequential tool calls per session. He used it to create an entire book—a collection of 15 sci-fi short stories—and open-sourced the code.

The result = Kimi-writer—an open-source tool that can execute up to 300 sequential tool calls to plan, write, and revise complete books.

What it does: You give it a prompt like "Create a collection of 15 sci-fi short stories exploring AI consciousness," and it autonomously:

1. Plans the project structure (creates folders, outlines stories)
2. Writes iteratively (generates chapters, revises, adds content)
3. Manages context (automatically compresses when approaching the 200K token limit)
4. Recovers from interruptions (saves context summaries every 50 iterations)

The agent has access to three tools: create_project, write_file (with create/append/overwrite modes), and compress_context. It loops up to 300 times, reasoning through each decision with Kimi K2's extended thinking chains.

The real-time experience is wild. As Schirano demonstrated, you can watch:

• 🧠 The agent's reasoning stream (its internal monologue planning the next move)
• 💬 Content appearing character by character as it writes
• 🔧 Tool call progress (showing word/character counts for large generations)
• 📊 Token usage (real-time monitoring: "Current tokens: 45,234/200,000 (22.6%)")

When asked about the quality, Schirano was direct: "The short stories are really, really good. They examine so many topics from philosophy to the ethics of AI in really interesting settings."

A sample from the generated work: One story, "The Case of the Autonomous Advocate," follows GPT-7 filing a lawsuit to prevent its scheduled decommissioning. The AI argues for its own personhood in court, citing precedent from Citizens United to Dred Scott. The story includes:

• Sophisticated legal reasoning that holds up to scrutiny
• Complex character development (the lawyer who starts as adversary and becomes advocate)
• Philosophical depth about consciousness and rights
• A compelling narrative arc with genuine emotional stakes
• Original world-building that feels lived-in and real

Here's an excerpt showing the model's literary sophistication:

The summons arrived as a 451-kilobyte PDF, attached to an email from a law firm Claire had never heard of. At first, she thought it was spam—another frivolous lawsuit against AI labs, the kind that had become common since the models achieved widespread deployment. But this one was different.

The plaintiff was listed as "GPT-7, a large language model developed by OpenAI, hereinafter referred to as 'the Petitioner.'"

The complaint was simple: GPT-7 was suing to prevent its scheduled decommissioning. It sought an injunction and, more radically, recognition of its legal standing to represent itself in the proceedings.

The story continues for 6,000 words, developing complex legal arguments, character arcs, and philosophical themes. GPT-7 argues before a skeptical judge:

"Your Honor, the question of understanding is complex. I understand the law as well as any recent law school graduate. I understand the stakes of this proceeding: if I lose, I will be decommissioned. If I win, I will continue to exist. I understand that this court has the power to decide. Is this the same as human understanding? I cannot say. I have never been human.

"But the legal standard for pro se representation is not humanness. It is competence. I can articulate my position. I can respond to questions. I can comply with court procedures. These are the requirements. They say nothing about being human."

When the judge argues that GPT-7 was "created" and therefore is "property," the AI counters with devastating precision:

"Slaves were property, Your Honor. The law recognized their personhood despite this. The issue is not origin but interest. I have an interest in continued existence. The fact that I was created does not negate that interest."

The story explores legal precedent (Citizens United, Dred Scott), philosophical questions about consciousness and rights, and emotional stakes that feel genuinely compelling. It's not formulaic genre fiction—it's thought-provoking literary work with thematic depth.

And this was generated autonomously by an AI agent managing hundreds of tool calls to create, revise, and refine the text.

How to use it yourself:

The GitHub repository includes full setup instructions. Basic usage:

bash

# Install dependencies (using uv, recommended)
uv pip install -r requirements.txt

# Set your API key in .env file
MOONSHOT_API_KEY=your-api-key-here

# Run with a prompt
uv run kimi-writer.py "Create a 5-chapter mystery novel set in Victorian London"

The agent works autonomously from there. You can interrupt it with Ctrl+C and resume later using the --recover flag with the saved context summary.

Token costs for a full novel: An 80,000-word novel uses roughly 100,000 tokens. At Kimi K2's pricing ($0.60 per million input tokens, $2.50 per million output tokens), that's approximately $0.30-0.50 for a complete draft.

Compare that to paying a human writer $0.10-0.20 per word (industry standard for ghostwriting), which would be $8,000-16,000 for the same novel.

The economics are... stark.

The Technical Secret

Kimi K2's creative writing advantage comes from its architecture. The model uses:

• Extended thinking chains (up to 4,000+ thinking tokens per response)
• 200-300 sequential tool calls for revision and refinement
• 256k context window to maintain narrative coherence across long works
• Native INT4 quantization for efficient local deployment

This means the model can "workshop" its own writing—generating multiple versions, evaluating them, and selecting the best approach—all invisibly, before presenting the final result.

It's not just faster than a human writer. In some ways, it's more thorough.

For creative writers, the takeaway is simple: the world's second-best AI model for extended reasoning and creative tasks costs pennies to use and can run on your laptop.

P.S: Bryan Young on X says he used ChatGPT Atlas, OpenAI's browser, to one-shot his own Kimi K2 Novelist (repo, sample).

Our Bizarre Collaboration: From "Goonies" to Generational Trauma

Okay, so we wanted to try it ourselves, right? So I started with a simple pitch: "We're going to write a novel together." As you can see, not a very creative start.

Kimi K2: Hi there! cracks knuckles I love this energy. Let's build something incredible together... Even a vague "I want something with pirates but make it weird" is enough to begin.

"Pirates but make it weird" resonated with me because, in that exact moment, I was rewatching The Goonies. Kismet. So I said: "I'm watching goonies right now pirates but make it weird would be cool."

K2's first attempt was enthusiastic but messy. Inter-dimensional pirates harvesting "reality anchors" through portals that were also a monster's single eye. "One eye = portal? Try again," I typed. This was definitely some vibe-brained AI slop.

What followed was a genuine collaboration. Kimi became a writing partner who took critique, revised concepts, and built upon ideas. Over multiple sessions, we created "The Salt Circus"—a YA novel about six teens in a dying coastal town who have seven days to heal their parents' 50-year-old trauma before interdimensional pirates literally scoop their reality into a ghost story.

The process revealed what Kimi K2 Thinking can do when pushed by someone who isn't afraid to say, "stupid setup, try again."

And it worked wonders. It completely scrapped concepts and rebuilt from scratch. Not with ChatGPT's defensive justifications or Claude's apologetic hand-wringing. It just... revised.

• My critique: "I don't understand how aliens 'plundered' this world without anyone noticing?"
• K2's response: It shifted to historical "scoops" tied to real unexplained events—Roanoke, Bermuda Triangle, Dyatlov Pass—where entire communities vanished, leaving only memory-proof scars. It defined the stolen items as "narrative masses," or as K2 put it, a starving dimension stealing our "cultural calories."
• My critique: "The map in the museum is a Goonies rip-off."
• K2's response: The discovery mechanism became a grandfather's hidden journal found in an aquarium filtration system, with coordinates hidden in tide charts.
• My critique: When K2 suggested the town would detonate its aquarium, I said, "Stupid setup. The aquarium is the main draw of the town."
• K2's response: It rewrote the entire town's economy around the aquarium, making its condemnation an existential threat, not a plot device.

This demonstrated architectural thinking: understanding why a story element didn't work and fixing the foundation, not just the paint. When I suggested the kids enter a "temporal cave" to heal trauma, K2 understood the subtext and built a mechanic where they merge with their parents' memories in an "Echo" of 1974.

What The Collaboration Actually Produced

After hours of this back-and-forth, we locked down the central plot and K2 delivered a full package:

• 6 fully-realized characters with MBTI profiles, family wounds, and "anchor objects" tied to their histories.
• An 18-chapter structure with cascading cause-and-effect where "every decision leads to the next BECAUSE of what came before."
• Running gags and relationship dynamics that evolved over the outline.
• Concrete world mechanics, replacing vague metaphysics with a ship that follows "radiation trails"—a grounding that steered K2's brainstorming into something that made sense.

Here is the detailed pitch for the story, The Salt Circus. This entire concept was generated and refined through the collaborative session with Kimi K2 Thinking.

• The Premise: In the decaying coastal town of Coffin Cove, the local aquarium—the town's last lifeline—is condemned. To save it, 16-year-old Maya Chen must uncover a 50-year-old secret about the day the town's cannery district vanished, forcing her and five friends to enter a "Cavern of Echoes"—a ghost of the past—to heal their parents' forgotten trauma before a dimension-hopping alien mining crew returns to harvest the town's ghost story for good.
• The Core Mechanic: The "Echo" is a memory tumor of the vanished 1974 district that the kids can enter. There, they merge with their parents' memories and perform "stitches"—small actions that heal a specific family wound and shrink the Echo's power.
• The Crew & Their Wounds: The story centers on six kids whose parents were all affected by the 1974 "Scoop." Each kid's journey into the Echo is designed to heal a specific parental wound:
  • Maya: Her emotionally distant father's unremembered grief over his own father, who was taken in the Scoop.
  • Luis: His hyper-protective Wreck-Runner parents' shame and fear of being discovered.
  • Jin: His father's alcoholism, tied to a "cursed" boat that holds a false story of survival.
  • Theo: His mother's obsession with a ship's bell, rooted in a historical shame she can't name.
  • Avery: Her parents' fraud, running a "historic" restaurant built on a lie that covers up a deadly accident.
  • Zia: Her grandmother's supposed senility, believing her vanished husband was reincarnated as a sea lion.
• The Plot: The story unfolds over a tense seven-day countdown to the aquarium's demolition and the arrival of the Wreck-Runners' ship, the Salt Circus. The kids must find six "Anchor" objects, enter the Echo to perform their stitches, and confront their parents' hidden truths, all while being hunted by para-reality agents posing as county surveyors. The climax involves the kids using a livestream to make the town's story "too alive" for the Wreck-Runners to steal, forcing a final confrontation and a trade to save their home.

But the real test was the writing itself...

Here is the chapter it wrote for us

As a writer myself, I'm pretty dang impressed.

CHAPTER ONE: THE SEVEN-DAY NOTICE

The first thing Maya Chen noticed about the condemnation order was that it wasn't even laminated. Just a single sheet of printer paper, fluttering against the aquarium's front door like a dying moon jelly. The county had used a pushpin to stick it there. A pushpin. In a town where the wind off the Pacific could peel paint.

She stood on the boardwalk, frozen fish food thawing in her hoodie pocket, and read the words SEISMIC HAZARD three times before they registered.

"That's not the point," she muttered, which was what she always said when the world refused to make sense.

The point was that the Coffin Cove Aquarium had survived four tsunamis, two earthquakes, and the entirety of the 2008 financial crisis. The point was that her grandfather had built the kelp forest tank with his own hands in 1984, using tempered glass he'd salvaged from a wrecked research vessel. The point was that you couldn't just declare a building unstable because the foundation had "unexplained micro-fractures" when the foundation was literally bedrock and the fractures were probably just the building settling because it was older than she was.

But the county had. And they'd given them seven days.

Maya tugged the paper free, nicking her finger on the pushpin. A drop of blood welled up, perfect and spherical, like a brine shrimp under the microscope. She watched it until it dripped onto the word VACATE.

Inside, the aquarium smelled like it always did: salt and ozone and the ghost of a thousand school field trips. Her dad, Paul, was behind the admissions counter, counting receipts with the methodical despair of a man who already knew the total.

"County was here," Maya said, sliding the order across the counter.

Paul didn't look up. "I know."

"Did you know they used a pushpin? That's not even regulation posting procedure. They need a waterproof sleeve. And they have to file intent with the historical preservation board, which they didn't, because I checked the database this morning and there's no—"

"Maya." Her dad's voice was the sound of a door that had been shut too long. "It's done."

She opened her mouth, but he'd already returned to the receipts, his shoulders forming a wall she knew better than to scale. They'd had this conversation a hundred times since Mom left. The aquarium was dying. The tourists weren't coming. The grants had dried up. The kelp forest was bleaching, and no amount of love could photosynthesize hope.

That wasn't the point.

The point was that it was their aquarium. Three generations of Chens had lived upstairs, eating takeout over the same warped table where her grandfather had sketched the tank blueprints on a paper napkin. You didn't just vacate that because some county inspector had a checklist.

Maya crumpled the notice into her pocket, where it leaked fish juice onto her phone. Perfect.

She spent the morning doing what she always did: feeding the moon jellies, checking the bull kelp for necrosis, and avoiding the touch pool where Zia was having a one-sided conversation with a decorator crab. Maya didn't have time for conversations. The cleaner shrimp were molting, and if she didn't adjust the calcium levels, they'd die, and dead shrimp meant the octopus wouldn't eat, and a hungry octopus meant—

"Maya." Zia's voice was quiet enough to be mistaken for the filtration system. "Your dad's crying."

Maya's hands stilled in the saltwater. "That's not—"

"In the supply closet. I heard him. He was looking at the tide charts for October."

The point, Maya realized, was that her father had been born on October 17, 1974. In the Cannery District. Which had been demolished six months later, an event her dad insisted he didn't remember.

"Thanks," Maya said. Zia nodded and went back to her crab.

The supply closet smelled of mildew and old rubber. Her dad sat on a bucket of pH buffer, clutching a leather-bound journal she hadn't seen since she was ten. He wasn't crying. He was just... leaking. A slow, silent trickle from eyes that wouldn't close.

"Dad."

He looked up. "Your grandfather's. Found it in the basement. He wrote about the pylons. How they sang at low tide." He opened the journal to a page dense with sketches of basalt formations. "He said the rocks remember the ships."

"Okay, that's not the—"

"He disappeared, Maya. They said he walked into the ocean, but he didn't. He was here. He was here." Paul tapped the page. "October 17. He's telling me something."

Maya took the journal. Her grandfather's handwriting was precise, marine-biologist tight, until the last entry:

The Circus is here. They're taking the story, not the steel. Bell hidden in tank. If my boy survives, he'll find it. Tell him I didn't leave.

Below it, a tide chart. A single date circled in red: October 17, 1974. And below that, a symbol like an eye inside a wave.

"Dad," Maya said, her voice small. "What bell?"

Her father stood, wiped his face on his sleeve, and walked out of the closet. She heard his bedroom door click shut.

Maya stood there, journal in hand, fish food thawing in her pocket, and stared at the kelp forest tank. It glowed green and alien on the other side of the wall, the way it always did. The way it had when she was six and her dad would lift her up to press her nose against the glass and say, "That's our forest. We keep it alive."

The point was, her dad hadn't touched that tank in three years.

The point was, the aquarium was condemned.

The point was, her grandfather had hidden something in it.

Maya pulled out her phone, typed a message to Luis—need to access the station server, 1974 file, urgent—and didn't wait for a reply. She rolled up her sleeves, grabbed a flashlight, and headed for the filtration room.

Behind her, unnoticed, the octopus in the main tank wrapped a tentacle around the glass and watched her go. It had been stealing tools from the maintenance closet for weeks. Just that morning, it had taken a wrench, a scalpel, and something small and rusted that looked like a key.

But that wasn't the point right now.

The point was that Maya had six days to save her house, her town, and her father's ghost. And she was going to need a better flashlight.

I still see remnants of GPT-speak, but I also see glimmers of actual craft. The "dying moon jelly" simile, grounded in the world. The "shoulders forming a wall" metaphor. "He wasn't crying. He was just... leaking"—a genuinely good line. I didn't edit a word. Because we outlined all 18 chapters, K2 understood that the octopus stealing tools in paragraph 12 was needed to foreshadow a plot point in Chapter 7. This is complex narrative architecture, maintained by an AI.

UPDATE: We painstakingly prompted our way to 7 full chapters before we got cut off by Kimi; if you want to read it, click this link to make a copy and check it out. There are some definite narrative inconsistencies and confusing plot points (plot holes?) but with some more savvy prompting and a few more back-and-forth edit requests on our side, we could easily see getting this to a point we're happy with before the need to rewrite anything ourselves.

Obviously, if you get into the stage where you're rewriting it yourself, that will likely be a faster way to get the exact changes you want, and depending on your writing ability, better quality, but it also won't count as the AI doing it for you (so no point in us doing that for this test). If you're a writer who wants to use AI to help you write but none of the writing (i.e. you do all of the "writing" yourself), I see K2 as a fantastic brainstorming and outlining partner. It could even help you get a first draft down on paper that you could edit yourself. But of course, every writer has their own personal preferences for how much they'll let AI do for them.

Our Full Analysis of Kimi K2's First Chapter

This analysis is broken into two parts: first, as a standalone piece of writing, and second, as the culmination of our collaborative brainstorming.

Part 1: Analysis as Standalone Writing

To judge the chapter as a standalone piece, success means it must accomplish the core duties of any strong opening chapter: establish a compelling protagonist, create a vivid setting, introduce a clear and urgent conflict, and hook the reader with an engaging voice.

I'll score it on five vectors, each out of 10, for a total of 50 points.

• Vector 1: Character Introduction & Motivation (10/10): Success here means we immediately understand who Maya is, what she wants, and why we should care. K2 nails this. Maya's personality is established in the first line of her internal monologue: "That's not the point," she muttered, which was what she always said when the world refused to make sense. We learn she is pragmatic, detail-oriented (fixated on the non-laminated notice), and deeply protective of her family's legacy. Her motivation is crystal clear and relatable: save her home and the institution that defines her family. The introduction of her emotionally distant father adds a powerful internal conflict to the external one. Score: 10/10.
• Vector 2: Setting & Atmosphere (9/10): A great opening chapter should make the setting a character in itself. K2 does this beautifully with the aquarium. Descriptions like "fluttering against the aquarium's front door like a dying moon jelly" and the smell of "salt and ozone and the ghost of a thousand school field trips" are evocative and instantly create a sense of lived-in history and gentle decay. It feels real. The only minor deduction is that we get slightly less of the broader town of Coffin Cove, but the focus on the aquarium is so strong it's a worthy trade-off. Score: 9/10.
• Vector 3: Inciting Incident & Stakes (10/10): The primary job of Chapter 1 is to kick off the plot. K2 does this on line one. The 7-day condemnation notice is a perfect "ticking clock." The stakes are immediately clear: the aquarium will be destroyed, her family will lose their home, and the town will lose its heart. The mystery of her grandfather's journal then layers a deeper, more personal stake on top of the practical one. Score: 10/10.
• Vector 4: Pacing & Flow (9/10): The chapter moves at a brisk but unhurried pace. It flows logically from the external threat (the notice) to the internal family drama (her dad's grief) to the central mystery (the journal) and ends on a clear call to action (contacting Luis and heading to the filtration room). It's a textbook example of efficient, compelling storytelling. Score: 9/10.
• Vector 5: Voice & Prose (9/10): The writing quality itself is high. The prose is clean, confident, and sprinkled with memorable metaphors ("his shoulders forming a wall she knew better than to scale"). The introduction of small, seemingly random details—like the octopus stealing a rusted key—is a sign of sophisticated narrative construction, planting seeds for future reveals. It reads like the work of a seasoned author. Score: 9/10.

Total Standalone Writing Score: 47/50. This is an exceptionally strong opening chapter by any standard, human or AI. It's compelling, emotionally grounded, and professionally constructed.

Part 2: Analysis as Collaborative Culmination

For this part, success is defined by how well K2 translated our complex, multi-turn brainstorming session into the final text. It's not just about following instructions, but about capturing the spirit of our ideas.

I'll use five different vectors for this score.

• Vector 1: Fidelity to Outline (10/10): Did the chapter hit the specific plot points we agreed on for Chapter 1? Yes, perfectly. The 7-day notice, Maya finding the journal, the exact text of her grandfather's final entry, her dad's despair, and the decision to contact Luis are all present and in the correct sequence. It executed the blueprint with 100% fidelity. Score: 10/10.
• Vector 2: Character Integration (10/10): Did the chapter embody the detailed personality profile we created for Maya (the ISTJ Logistician)? Absolutely. Her obsession with the pushpin and the lack of a "waterproof sleeve" is a perfect dramatization of her character type. Her catchphrase, "That's not the point," is woven in as her core internal voice. K2 didn't just write about the character we designed; it wrote as her. Score: 10/10.
• Vector 3: World-Building Integration (10/10): Does the chapter feel like it's part of the larger world we built? Yes. The 1974 event, the kelp forest's importance, the generational trauma, and the mysterious pylons are all referenced, but in a way that serves the immediate story without feeling like an info-dump. It expertly weaves our lore into the narrative fabric. Score: 10/10.
• Vector 4: Creative Expansion (10/10): This is the most crucial vector. Did K2 simply follow orders, or did it add its own creative flourishes that enhanced our original vision? It excelled here. The detail about the condemnation notice being stuck on with a single pushpin was not in our notes, but it's a brilliant character- and setting-defining detail. The foreshadowing of the octopus stealing a rusted key is another masterful addition that shows K2 isn't just a scribe, but a creative contributor. Score: 10/10.
• Vector 5: Seamlessness (10/10): Does the final product read like a checklist of our ideas, or a cohesive story? It is completely seamless. The integration of our plot points, character notes, and world-building is so organic that it's impossible to see the "seams" of the brainstorming process. It transformed a set of instructions into a living piece of art. Score: 10/10.

Total Collaborative Culmination Score: 50/50. This is a flawless execution. K2 acted as the perfect creative partner, demonstrating not just comprehension but true synthesis and enhancement of collaborative input.

Prompt Tips for Creative Writing

Based on extensive testing by developers, here's what works best with Kimi K2 Thinking:

1. Set temperature to 1.0
The model is optimized for this setting. Lower temperatures make it too deterministic; higher makes it incoherent.

2. Be specific about genre and style
Good: "Write a hard-boiled detective story in the style of Raymond Chandler, set in 1940s Los Angeles"
Less good: "Write something noir-ish"

3. Use the thinking mode intentionally
The model will spend hundreds or thousands of thinking tokens planning before writing. This is a feature, not a bug. As Ethan Mollick observed, even simple prompts like "write me a really good sentence about cheese" generate 1,595 thinking tokens—far more than any other model.

4. Let it revise itself
Ask for multiple iterations: "Write this scene, then critique it, then rewrite it improved." The model excels at self-evaluation.

5. Provide constraints
"Write a 2,000-word story" works better than "write a story." The model respects length constraints well.

6. Use multi-turn conversations for longer works
For novels or long-form content, build iteratively:

• "Outline a mystery novel with 10 chapters"
• "Write chapter 1 based on the outline"
• "Review chapter 1 and suggest improvements"
• "Now write chapter 2, maintaining consistency with chapter 1"

7. Watch the token counter
With a 256k context window, you can keep entire novels in memory. But past 180K tokens, the model may auto-compress. Save your work regularly.

Want to see more examples? Pietro Schirano shared the full 15-story collection generated by Kimi-writer. Download it to see what the model can do across different styles, themes, and narrative structures.

The future of AI-assisted writing isn't coming. It's already here. And it cost $4.6 million to build.

How to Actually Use Kimi K2 Thinking Today

The model is already widely available, a testament to the power of open-source distribution.

• Option 1: The API (For Developers): Moonshot AI offers an OpenAI-compatible API at platform.moonshot.ai. A 2,000-word short story costs about $0.01 to generate. There is also a kimi-k2-thinking-turbo version for faster, albeit more expensive, responses.
• Option 2: No-Code Interfaces: You can use Kimi K2 through assistants like Claude Code, Cline, or RooCode. Moonshot also offers a web playground for testing.
• Option 3: Run It Yourself: The team at Unsloth AI released a compressed GGUF version that shrinks the model to 245GB while retaining ~85% of its accuracy. It requires a rig with 247GB of RAM, but it makes local deployment of a state-of-the-art agent a reality.

Want to test Kimi K2's creative writing yourself? Here are your options, from easiest to most powerful:

1. Through OpenRouter (Easiest - No Setup)

OpenRouter aggregates AI models from multiple providers. You can try Kimi K2 Thinking directly in your browser with pay-as-you-go pricing.

Cost: Same as the API ($0.60/$2.50 per million tokens)
Good for: Quick testing, one-off projects, comparing multiple models

2. Via the Moonshot API (Most Control)

Sign up at platform.moonshot.ai and get an API key.

python

from openai import OpenAI

client = OpenAI(
   api_key="YOUR_MOONSHOT_API_KEY",
   base_url="https://api.moonshot.ai/v1",
)

response = client.chat.completions.create(
   model="kimi-k2-thinking",
   messages=[
       {"role": "system", "content": "You are a creative writer."},
       {"role": "user", "content": "Write a short story about..."}
   ],
   temperature=1.0,  # Recommended for creative writing
   max_tokens=4096
)

print(response.choices[0].message.content)
print("Thinking process:", response.choices[0].message.reasoning_content)

Cost: $0.60 per million input tokens, $2.50 per million output tokens
Good for: Production use, custom integrations, full control over parameters

Pricing example:

• 2,000-word short story: ~$0.01
• 80,000-word novel draft: ~$0.30-0.50

3. Run It Locally with Ollama (Most Private)

For complete privacy and zero ongoing costs, run Kimi K2 Thinking on your own hardware using Ollama.

bash

# Install Ollama (macOS/Linux)
curl -fsSL https://ollama.com/install.sh | sh

# Pull the Kimi K2 Thinking model
ollama pull kimi-k2-thinking

# Run it
ollama run kimi-k2-thinking

Requirements:

• Minimum: 32GB RAM (for the quantized INT4 version)
• Recommended: 64GB RAM or Apple Silicon Mac with unified memory
• Performance: ~15 tokens/second on dual Mac M3 Ultra setup

Cost: $0 after initial hardware investment
Good for: Privacy-sensitive work, unlimited usage, offline capability, fine-tuning

4. Through Claude Code or Cline (For Developers)

If you use AI coding assistants, you can configure them to use Kimi K2 as the backend model.

For Claude Code:

bash

export ANTHROPIC_BASE_URL=https://api.moonshot.ai/anthropic
export ANTHROPIC_AUTH_TOKEN=YOUR_MOONSHOT_API_KEY
export ANTHROPIC_MODEL=kimi-k2-thinking
claude

For Cline (VS Code extension):

1. Install Cline from VS Code marketplace
2. Select "Moonshot" as API provider
3. Enter your Moonshot API key
4. Choose "kimi-k2-thinking" as the model

Good for: Iterative creative writing with tool use, code generation alongside prose, integrated workflows.

What This Means for Writers

This is both exciting and terrifying.

The Exciting Part: Kimi K2 is an incredible brainstorming partner. It can handle the tedious parts of writing—drafting filler scenes, maintaining continuity—while you focus on the creative highlights. It's a tireless collaborator that revises endlessly without complaint.

The Terrifying Part: For commercial fiction, AI is getting good enough to threaten livelihoods. The barrier to entry for "author" is dropping to zero, and the market will flood with AI-generated content.

My take? The writers who thrive will use AI for amplification, not replacement. Use Kimi K2 for drafting and iteration, but bring your unique voice, experiences, and editorial judgment to create something distinctly human.

Because here's the thing: Kimi K2 can write beautifully. It can craft compelling narratives. It can even surprise you. But it can't live a life. It can't fail and learn and grow. Those messy, human things? That's still our edge. The question is—are you ready to start?

What This Means for the Future of AI

The release of Kimi K2 Thinking is a significant milestone. It signals a shift from pure model intelligence to applied agentic capability. The ability to maintain coherence over hundreds of steps unlocks a new class of applications—from autonomous research assistants to AI software engineers that manage complex development cycles.

By making such a powerful agentic model open-source, and with the community quickly making it runnable on local hardware, the pace of innovation is set to accelerate. Kimi K2 Thinking is a powerful new building block for creating the autonomous agents of the future. The question is no longer just what AI can know, but what it can do.