On July 27, 2026, a 1.56 TB file appeared on Hugging Face: the weights for Kimi K3, Moonshot AI's 2.8-trillion-parameter coding model. Hugging Face's CEO posted that K3 was the top trending model on the Hub within 30 minutes, the fastest release growth the platform had seen. That is a striking number for a model almost nobody can actually run on their own hardware. The gap between "the weights are out" and "the weights are usable" is where the real story lives.
What Kimi K3 actually scores - and how to read the numbers
Moonshot AI released Kimi K3 in mid-July 2026: a 2.8-trillion-parameter mixture-of-experts model with 104 billion parameters active per token, native visual understanding, and a 1 million-token context window.
K3 leads on Program Bench (77.8 vs 76.8), SWE Marathon (42.0 vs 35.0), BrowseComp (91.2 vs 88.0), and Frontend Code Arena (1,679 vs 1,631 Elo) compared to Claude Fable 5.
Fable 5 is stronger on deep reasoning; K3 is stronger on sustained execution and frontend code.
But the headline SWE-bench Verified number needs a footnote. K3's SWE-bench Verified score dropped from K2.6's 80.2% to 76.8%. This likely reflects different benchmark versions or evaluation conditions rather than an actual regression - K3's broader benchmark profile is clearly stronger than K2.6's.
K3 was evaluated with the Kimi Code harness, while other results come from frontierswe.com
- so when you see a table ranking K3 against Claude or GPT on SWE-bench, you are often comparing outputs from different agent scaffolds, different attempt budgets, and different problem sets. The score measures the system, not just the model.
The more useful signal is on benchmarks where all models were evaluated in the same harness. K3 scores 57 on the Artificial Analysis Intelligence Index, the top open-weight result, with GLM-5.2 next at 51.
| Benchmark | Kimi K3 | Claude Fable 5 | GLM-5.2 | DeepSeek V4-Pro |
|---|---|---|---|---|
| Terminal-Bench 2.1 | 88.3 | - | 81.0 | - |
| SWE Marathon | 42.0 | 35.0 | - | - |
| Program Bench | 77.8 | 76.8 | - | - |
| SWE-bench Verified | 76.8% | leads | - | 80.6% |
| SWE-bench Pro | - | - | 62.1% | - |
| AA Intelligence Index | 57 | above | 51 | - |
One independent real-world test cut through the noise cleanly. In one team's evaluation on their own production Rails codebase, Kimi K3 matched Opus 4.8's quality for about a quarter of the cost - the best any open-weight model had done on their repo - but took twice as long to get there. That time penalty matters for interactive coding but not much for overnight batch jobs or CI pipelines.
Can you actually self-host it? The hardware math
In 2026, Kimi K3 can be self-hosted, but the public weight repository is about 1.56 TB and the official vLLM recipe starts at 8 NVIDIA GB300 GPUs or 8 AMD MI355X/MI350X GPUs. A single 8xH100 node tops out at 640 GB of VRAM - less than half the model's footprint. This is not a one-pod experiment.
Kimi Delta Attention is not in llama.cpp, Ollama, or LM Studio as of late July 2026, and integrating it into serving frameworks like vLLM takes weeks to months of community work after a release.
Downloading the weights today does not mean your existing stack can serve them today. The model card officially names vLLM, SGLang, and Moonshot's own TokenSpeed as supported runtimes - but that means you need to run one of those, configured correctly, not reach for whichever tool you already have open.
The architecture introduces Kimi Delta Attention for 6.3x faster decoding and attention residuals that improve token efficiency by 25% at 2% compute cost. Those are Moonshot's numbers. The practical payoff is that Kimi Delta Attention cuts KV cache memory by up to 75% at 1M context versus a standard transformer, making the 1M window actually usable in production. For a model at this scale, that matters: a vanilla transformer at 1M tokens would need far more memory than the weights themselves.
For most teams, the realistic K3 architecture in August 2026 is the routed one: API access for K3, open-weight deployment for smaller models like GLM-5.2 or Qwen 3.6 where the hardware math works, and a gateway in front of all of it. A teammate like Beagle, routing requests from Slack to the right model endpoint, fits naturally into that gateway layer - you pick the model; the plumbing stays the same.
Beagle in action#eng-tools, 2:40pmThe license: not MIT, not a problem for most - but read clause 2
The most accurate label is open-weight under a custom, broadly permissive license - not unconditionally open source.
Moonshot's own framing is "open-weight," which is precise. The distinction has teeth for one specific group.
The Kimi K3 license grants broad rights to use, modify and deploy the model for commercial purposes, but requires companies operating a "Model as a Service" business with aggregate annual revenue exceeding $20 million to enter into a separate commercial agreement with Moonshot AI. Note: the wording refers to aggregate revenue across the licensee and its affiliates; it does not say the threshold is limited to revenue generated by K3. A company that makes $25M/year selling SaaS and adds a K3-powered feature may already be past the threshold, regardless of how much revenue that feature generates.
Additionally, any commercial product or service using the model that reaches more than 100 million monthly active users or generates more than $20 million in monthly revenue must prominently display "Kimi K3" in its user interface.
Internal use of the model - defined as any use that does not make the software, its outputs, or its underlying capabilities available to third parties - is exempt from both requirements.
A MaaS business past $20M in trailing revenue is Together, Fireworks, Baseten, or someone at that weight class. For the overwhelming majority of engineering teams building internal tools, the license is effectively permissive. But legal review before a commercial product launch is the right call, not an assumption.
Where K3 sits against the open-weight field
GLM-5.2 is the strongest MIT-licensed model at 62.1% SWE-bench Pro, DeepSeek-V4-Pro resolves 80.6% of SWE-bench Verified under MIT, and Qwen3.6-27B hits 77.2% SWE-bench Verified in a 27B dense package. That 27B dense number matters: Qwen3.6-27B fits on hardware most teams already have.
The comparison that is not being made often enough: K2.6 was priced at $0.60/$2.50 per million tokens, one of the cheapest frontier models. K3 at $3.00/$15.00 is competitive but no longer the budget option.
For existing K2.6 users, the upgrade decision hinges on whether the coding improvements and 1M context justify a 5-6× price increase. If your workload is primarily single-file edits and quick completions, K2.6 or K2.7 Code remain the rational choice. K3 earns its compute on sustained multi-file, multi-hour sessions - and on evaluations that specifically reward long-context coherence.
The differentiator isn't the weights anymore - it's how well you run them. K3 is the clearest proof of that so far. The model exists, it benchmarks at the frontier, the weights are downloadable. The work now is in inference infrastructure, evaluation harnesses that give you comparable numbers across models, and license compliance for anyone building on top of it commercially.
Kimi K3 open-weight model: common questions
Is Kimi K3 the best open-weight coding model?
On sustained coding benchmarks - SWE Marathon, Program Bench, Terminal-Bench 2.1 - yes, K3 currently leads the open-weight field. On single-call SWE-bench Verified, DeepSeek V4-Pro (80.6%) scores higher. The honest answer is that the winner depends on which benchmark you weight and which agent harness ran the numbers.
Can you run Kimi K3 locally with Ollama or llama.cpp?
Not yet, as of late July 2026. K3's custom Kimi Delta Attention architecture is not yet supported in llama.cpp, Ollama, or LM Studio. Self-hosting requires vLLM or SGLang on at least an 8x B200 GPU node - roughly 1.4-1.56 TB of VRAM. Community support for mainstream tooling is expected to develop over the following months.
Is Kimi K3 free for commercial use?
For most teams, yes. Internal use is explicitly exempt from all license requirements. Commercial products are free to use K3 unless you are operating a Model-as-a-Service business with aggregate annual revenue over $20 million, or building a consumer product with over 100M monthly users. Read the license in full before a commercial launch.
How does Kimi K3 compare to GLM-5.2 and DeepSeek V4?
K3 scores higher on benchmarks requiring long-context and sustained execution. GLM-5.2 is MIT-licensed and fits smaller infrastructure (117B total, 5.1B active per token). DeepSeek V4-Pro matches K3 on SWE-bench Verified and is also MIT-licensed. For teams that need a model to run on a single node, neither K3 nor GLM-5.2's full MoE is realistic - Qwen3.6-27B at 27B dense parameters is the practical alternative.
What does "open-weight" mean for Kimi K3 versus open source?
Open-weight means the trained weights are published and downloadable. Open source, under the OSI definition, also requires the training data and training code. K3 publishes weights under a custom license but does not release training data. Moonshot calls it open-weight - that label is accurate.