Kimi K3 Review: Moonshot's 2.8T Open Model vs Fable 5 and GPT-5.6 Sol

Kimi K3 Review: Moonshot's 2.8T Open Model vs Fable 5 and GPT-5.6 Sol

Last verified: July 17, 2026. Every spec, price, and benchmark below was re-checked against Moonshot's platform docs, Artificial Analysis, and independent coverage the day this went live. Kimi K3 launched July 16, 2026.

In this Kimi K3 review, one launch number does the talking: Moonshot's new model debuted at #1 on Arena's Frontend Code leaderboard, a 17-place jump over Kimi K2.6, ranked above Claude Fable 5. That's a Chinese open-weight model winning a front-end coding contest judged by real developers in blind matchups. Under the hood it's a 2.8-trillion-parameter Mixture-of-Experts design that fires only 16 of 896 experts per token, reads a million tokens of context, and prices input between $0.30 and $3.00 per million. The catch you need to know before you get excited: you can't download the weights yet, and Moonshot says that changes on July 27.

Quick verdict:

• What it is: Moonshot AI's flagship, a 2.8T-parameter MoE model with 1M context, native image and video input, and always-on reasoning. API id kimi-k3.
• Where it wins: agentic browsing (BrowseComp 91.2) and long-horizon coding (SWE Marathon 42.0, above both Fable 5 and GPT-5.6 Sol). #1 on Arena's Frontend Code Arena.
• Where it trails: FrontierSWE and HLE-Full (Fable 5 leads), DeepSWE (GPT-5.6 Sol leads). Independent Artificial Analysis ranks it #4 of 189 overall.
• What it costs: $0.30 cache-hit / $3.00 fresh input, $15.00 output per million tokens. The priciest model any Chinese lab has shipped.
• The catch: open-weight in name, but the checkpoint doesn't go public until July 27, 2026. Until then, no self-hosting and no independent third-party evaluations.

If you want the one-line answer: Kimi K3 is the first open-weight model that trades real blows with the closed frontier, but it's launch-day software with a pending weight release and a premium price. Read on for the specs, the benchmark reality (including a caveat that changes how you should read every number), the pricing math, and who should actually run it.

What Is Kimi K3?

Kimi K3 is Moonshot AI's newest large language model, released July 16, 2026. It's a Mixture-of-Experts model with 2.8 trillion total parameters that activates just 16 of its 896 experts on each token, so the compute cost per token stays far below what a dense 2.8T model would demand. It accepts text, images, and video, holds a one-million-token context window, and keeps reasoning switched on by default.

Here's the spec sheet at a glance.

The interesting engineering story is the attention design. Moonshot calls it Kimi Delta Attention, or KDA, a hybrid linear attention mechanism that the company reports delivers up to 6.3x faster decoding in million-token contexts. K3 pairs it with a stack of newer tricks the lab names Attention Residuals, Gated MLA, and Stable LatentMoE. You don't need to memorize the acronyms. What they add up to is a model that can stay fast while carrying an enormous context, which is exactly the workload that breaks older architectures.

Put K3 next to the model it replaces and the jump is large. It nearly triples the parameter count of Kimi K2 (2.8T versus roughly 1T), quadruples the context window of the K2.6 and K2.7 line (1M versus 256K), and adds image and video input to a family that used to be text-first. If you tried an earlier Kimi and shrugged, this is a different animal.

Kimi K3 Benchmarks: Where It Wins and Where It Doesn't

Kimi K3's launch benchmarks are genuinely strong, and genuinely mixed. It tops the field on some coding and agent tasks and clearly trails the closed frontier on others. Before the table, one caveat that matters more than any single score: Moonshot ran each model inside its own coding environment. K3 was scored in KimiCode, Fable 5 in Claude Code, and GPT-5.6 Sol in Codex. The software wrapped around a model affects its results, so read these as directional, not as a settled leaderboard.

Here are the headline coding and agent numbers from Moonshot's launch table.

Read across the rows and a pattern shows up. K3 wins the long, grindy work. On SWE Marathon, which measures sustained multi-hour engineering sessions, K3's 42.0 beats both GPT-5.6 Sol and Fable 5 by a clear margin. It edges out the field on Program Bench, and it leads on the agent side too, posting 91.2 on BrowseComp and 91.1 on OmniDocBench, where Moonshot also reports the top spot on Automation Bench.

Where does it lose? On DeepSWE, GPT-5.6 Sol pulls ahead at 73.0. On FrontierSWE, Fable 5 is well clear at 86.6, though it's worth flagging that K3's 81.2 still beats Sol's 71.3 there. Moonshot's own table concedes that K3 trails Fable 5 on FrontierSWE and HLE-Full, and trails GPT-5.6 Sol on DeepSWE. This is not a model that beats the frontier everywhere. It's a model that beats it somewhere, which no open-weight release had really managed before.

The independent read backs that up. Artificial Analysis scores K3 at 57 on its Intelligence Index and ranks it #4 out of 189 models, behind Claude Fable 5 and two GPT-5.6 Sol reasoning settings but ahead of Claude Opus 4.8. Its measured output speed is a fairly modest 62 tokens per second. On the human-preference side, K3's #1 finish on Arena's Frontend Code Arena is the standout, because that ranking comes from developers voting on real front-end output rather than a self-reported score.

Independent developer Simon Willison ran K3 through his pelican-on-a-bicycle SVG test on launch day. The model produced a solid result and wrote accurate alt text for its own image, which speaks well of its vision. He also flagged the cost surprise we'll get to in the pricing section, and reminded readers that a quick SVG test tells you nothing about agentic tool calling, which is where a model like this earns its keep. Fair caution.

Kimi K3 vs Fable 5, GPT-5.6 Sol, DeepSeek and Qwen

So where does K3 sit in the wider field? Two comparisons matter: against the closed frontier, and against its fellow Chinese open-weight models.

Against the frontier, the honest framing is "frontier-adjacent, not frontier-topping." Claude Fable 5 and GPT-5.6 Sol still lead the aggregate intelligence rankings, and Fable 5 keeps a real edge on the hardest reasoning and frontier-coding evaluations. What changed with K3 is that the gap narrowed to single-benchmark trades rather than a category gulf. For a downloadable model to lead SWE Marathon and top a blind front-end coding vote is new territory.

Against its open-weight peers, K3 is the new high-water mark on raw capability, but it isn't the obvious default for every job. If you're weighing the Chinese open-weight options as a group, our Qwen vs DeepSeek vs GLM comparison walks through how those three families split the market: Qwen for the lightest self-host footprint and the most permissive license, DeepSeek for the cheapest tokens, GLM for hands-on coding. K3 now sits above all of them on peak benchmarks, and above them on price too. It's the premium option in a category that built its reputation on being cheap.

For a broader field including the models that actually beat GPT-4-class quality, our best open-source LLM roundup for 2026 puts K3's claims in context. The short version: K3 raises the ceiling for open weights, but "open weights" and "cheap" have officially stopped being the same thing.

Kimi K3 Pricing and the Cache-Hit Math

Here's where K3 gets divisive. Moonshot prices it at $0.30 per million cache-hit input tokens, $3.00 per million fresh input tokens, and $15.00 per million output tokens, flat across the entire million-token context.

Line it up against the model it replaces and the shift is stark.

That's roughly a 3x jump on input and nearly 4x on output versus K2.6. Willison noted the $3/$15 rate lands in the same tier as Claude Sonnet. The Decoder called K3 a signal that the era of super-cheap Chinese AI is ending. If your whole reason for running a Chinese model was the price, K3 will make you look twice.

But the cache-hit rate is the number that decides your real bill, so let's do the math on a realistic agentic loop, using Moonshot's published rates. Say your coding agent reads a 200,000-token codebase context and writes 8,000 tokens of output, and it does that 20 times a day:

• Cache-miss (first cold read): 200,000 input tokens at $3.00/M is $0.60, plus 8,000 output tokens at $15.00/M is $0.12. That's about $0.72 per call, or $14.40 a day.
• Cache-hit (repeated context, the common case in a coding session): 200,000 input tokens at $0.30/M is $0.06, plus the same $0.12 output. That's about $0.18 per call, or $3.60 a day.

The cache economics cut the input bill by roughly ten times, from $0.60 to $0.06 per call. Moonshot reports above 90% cache hits in coding workloads, which is the scenario that makes K3 affordable rather than eye-watering. The lesson for your budget: K3 is expensive on cold, one-shot calls and reasonable inside a warm, context-heavy loop.

One more cost trap Willison surfaced. K3 exposes only a single "max" reasoning level today, and reasoning tokens bill at the output rate. His simple SVG test burned 13,241 reasoning tokens on top of 3,417 output tokens, so a trivial prompt cost about 25 cents. There's no cheap "low reasoning" mode yet, which means K3 overpays on easy questions. If cost control is your priority, our guides on LLM API pricing and how to reduce LLM API costs apply directly here: cache aggressively, route trivial calls to a cheaper model, and reserve K3 for the hard, long-context work where it earns the premium.

Open Weights: What "Open" Actually Means Here

This is the part the launch headlines skated over. Kimi K3 is announced as an open-weight model, and Moonshot has a real track record of shipping downloadable checkpoints. But as of July 17, 2026, the K3 weights are not public. The Moonshot Hugging Face organization still lists only K2-series checkpoints. The company says the full weights arrive July 27, 2026.

Why does that gap matter? Three practical reasons:

• No self-hosting yet. You cannot run K3 on your own hardware or a private cluster until the weights drop. For teams with data-residency or air-gap requirements, K3 is API-only for now, which defeats a big part of why you'd choose an open model. When the weights do land, our guides on the best tools to run LLMs locally and running an LLM locally will get you started, though a 2.8T-parameter model is cluster-class, not laptop-class.
• No independent evaluations yet. Every K3 benchmark you've seen is vendor-run, in Moonshot's own environment. Serious third-party numbers on things like HLE or agent-specific tasks can't exist until outside labs have the weights to test. Treat the launch table as a strong claim, not a verified result.
• License terms are still settling. Prior Kimi releases used permissive licenses, but confirm the exact K3 license against the official model card when the checkpoint publishes, especially if you plan commercial deployment.

None of this makes K3 less impressive. It does mean that if your plan depends on downloading and fine-tuning the model, your real evaluation starts on July 27, not July 16.

How We're Evaluating Kimi K3 for Client Work

A quick note on where this review comes from, and where it stops. At Techsy, we wire third-party LLMs into production pipelines for B2B clients, usually through OpenAI-SDK-compatible endpoints so we can swap models without rebuilding the plumbing. K3 fits that path cleanly: it exposes an OpenAI-compatible API with the model id kimi-k3, so dropping it into an existing integration to trial it is a config change, not a rewrite.

Every time a new model lands, we run it through the same fixed evaluation on client-representative tasks before we recommend anything. And here's the honest limit on this review: we are not publishing our own K3 score yet. The weights aren't out, the launch benchmarks were vendor-run in Moonshot's own environment, and a fair number needs a neutral setup on tasks that look like real client work, not a leaderboard. Quoting a first-party benchmark from a one-day-old, weights-pending model would be exactly the kind of number we tell clients to distrust.

What we can evaluate today from the live API is the part that doesn't need a leaderboard: the integration surface, the cost model, and the failure modes. The pricing math above is our own calculation from Moonshot's published rates, not a benchmark, and it already tells you the operational truth: cache discipline is the difference between K3 being affordable and being a budget problem. If you want help figuring out whether a model like K3 belongs in your stack, that's the kind of evaluation we do at Techsy's AI integration practice, and you can book a free consultation to talk through it.

Who Should Use Kimi K3 (and Who Shouldn't)

Kimi K3 is a strong fit for a specific set of jobs and a poor fit for others. Match it to your actual workload.

Reach for K3 if:

• You run agentic browsing or research. Its BrowseComp and Automation Bench leads, plus the top independent agent-research read, make it the most capable open-weight option for tool-using agents right now.
• You do long-horizon coding. SWE Marathon is the benchmark that mirrors multi-hour engineering sessions, and K3 leads it. If your agents work across large codebases over long runs, this is its home turf.
• You want an open-weight frontier-adjacent model and can wait for the weights. If self-hosting a top-tier open model on July 27 is the goal, K3 is the most capable candidate available.

Hold off if:

• You need the weights today. Until July 27, K3 is API-only. An already-downloadable model serves you better this week.
• You run high-volume, cost-sensitive traffic. With a single expensive reasoning level and premium output pricing, K3 overpays on simple calls. Kimi K2.7-Code or a cheaper open model wins on bulk work.
• You require proven, independent evaluations before adoption. The launch numbers are vendor-run. If your process demands third-party verification, give it a few weeks.

Frequently Asked Questions

What is Kimi K3?

Kimi K3 is Moonshot AI's flagship large language model, released July 16, 2026. It's a 2.8-trillion-parameter Mixture-of-Experts model that activates 16 of 896 experts per token, supports a 1M-token context window, and accepts text, image, and video input with reasoning always on.

Is Kimi K3 open source?

Kimi K3 is announced as an open-weight model, but the weights are not public as of July 17, 2026. Moonshot says the full checkpoint will be released July 27, 2026. Until then it's available only through the Kimi apps and the API, so you cannot self-host it yet.

How is Kimi K3 different from Kimi K2?

K3 nearly triples K2's parameter count (2.8 trillion versus roughly 1 trillion), quadruples the context window of the K2.6 and K2.7 line (1M versus 256K tokens), and adds native image and video input to a family that was previously text-focused. It also introduces the new Kimi Delta Attention design.

How much does Kimi K3 cost?

Moonshot prices K3 at $0.30 per million cache-hit input tokens, $3.00 per million fresh input tokens, and $15.00 per million output tokens. That's roughly triple K2.6's input price and nearly quadruple its output price, making K3 the most expensive model a Chinese lab has released.

What is Kimi K3's context window?

Kimi K3 supports a context window of one million tokens, and pricing stays flat across that full window. The model uses a new attention design called Kimi Delta Attention, which Moonshot reports gives up to 6.3x faster decoding at million-token context lengths.

Can I run Kimi K3 locally?

Not yet. The weights aren't public until July 27, 2026. After that, self-hosting is technically possible, but a 2.8-trillion-parameter model needs cluster-class hardware rather than a single workstation, so most teams will still access it through the API.

Is Kimi K3 really better than Fable 5?

It depends on the task. K3 beats Claude Fable 5 on SWE Marathon, Program Bench, and Arena's blind front-end coding vote. Fable 5 still leads on FrontierSWE, HLE-Full, and overall aggregate intelligence rankings. Every launch benchmark also ran in each vendor's own environment, so treat single-benchmark wins as directional.

Is Kimi K3 good for coding?

Yes, especially for long, sustained coding sessions and front-end work, where it currently leads. It's strong on agentic and terminal tasks too. The main drawback is cost: with one expensive reasoning level, it overpays on trivial calls, so pair it with cheaper models for high-volume routine work.

How do I access Kimi K3?

You can use Kimi K3 through the Kimi web app, Kimi Work, and Kimi Code, or through the API with the model id kimi-k3. The API is OpenAI-SDK compatible, so adding it to an existing OpenAI-style integration is mostly a configuration change.

Key Takeaways

• Kimi K3 is the first open-weight model to genuinely trade blows with the closed frontier, leading SWE Marathon and topping Arena's blind front-end coding vote above Claude Fable 5.
• It's frontier-adjacent, not frontier-topping. Independent Artificial Analysis ranks it #4 of 189, and it trails Fable 5 on the hardest reasoning and frontier-coding tests.
• Open in name, pending in practice. The weights don't publish until July 27, 2026, so there's no self-hosting and no independent evaluation until then.
• The price ended the cheap-Chinese-AI era. At $3/$15 per million tokens, cache discipline is what keeps K3 affordable; budget for a warm, context-heavy loop, not cold one-shot calls.
• Best for agentic research and long-horizon coding. If you need weights today or run cost-sensitive high-volume traffic, wait or pick a cheaper model. Talk to us if you want help deciding.