Moonshot releases Kimi K3 open weights with 2.8T-parameter MoE

• Moonshot released Kimi K3 weights, report, and blog for a 2.8T MoE with 104B active parameters, native vision, and a 1M-token context window, according to Moonshot's release post.
• The license is open-weight, not clean-room open source: large model-as-a-service businesses need a separate agreement, and very large products must display the Kimi K3 name, per natolambert's license read and petergostev's summary.
• K3 is the new leading open-weight model on Artificial Analysis at 57, with only Anthropic and OpenAI models scoring higher in that index, according to ArtificialAnlys's index note.
• The release also shipped serving infrastructure, including FlashKDA, MoonEP, and AgentENV, while lmsysorg's SGLang post claimed 423 tok/s day-zero inference.
• The technical report leaves major cost inputs opaque: training tokens, total compute, and hardware details are missing in teortaxesTex's read and cedric_chee's infra note.

The Hugging Face model card has the dense spec table: 69 KDA layers, 24 Gated MLA layers, 896 experts, 16 active per token, MXFP4 weights, and MXFP8 activations. Simon Willison's note sizes the weights at 1.56TB and calls out the license shift away from K2's modified MIT language. The Moonshot release writeup says the open drop included MoonEP, FlashKDA, and AgentEnv, which makes the release feel less like a model card and more like a partial frontier-model factory manual.

Moonshot published three K3 artifacts in one drop:

• Model weights
• Technical report
• Technical blog

The Hugging Face model card lists the model as a 2.8T-parameter MoE with 104B activated parameters, 93 layers, 1 dense layer, 896 experts, 16 selected experts per token, a 160K vocabulary, and a 1,048,576-token context length.

K3 also has a usage gotcha: the GitHub README says thinking is always enabled, reasoning_effort supports low, high, and max, and multi-turn calls must pass back the complete assistant message, including reasoning_content and tool_calls.

The Kimi K3 License broadly allows use, modification, fine-tuning, deployment, and commercial products, but it adds two large-scale conditions.

• Model-as-a-service businesses with more than $20M in revenue over any consecutive 12 months need a separate Moonshot agreement.
• Commercial products with more than 100M monthly active users or $20M in monthly revenue must prominently display “Kimi K3.”

natolambert's clarification separates the two thresholds: $20M per year applies to model-as-a-service inference, while $20M per month or 100M users applies to products. Simon Willison notes that Moonshot describes K3 as open-weight rather than open source.

The model shape is the headline, but the useful part is the exact sparsity recipe.

• Total parameters: 2.8T.
• Active parameters: 104B.
• Experts: 896 total, 16 active per token.
• Attention: 69 KDA layers and 24 Gated MLA layers, a 3:1 mix in the model card.
• Precision: MXFP4 weights and MXFP8 activations through quantization-aware training.
• Context: 1M tokens.
• Vision: MoonViT-V2 with 401M parameters.

rohanpaul_ai's report summary pulls out the architectural bet: most layers no longer use full softmax attention, NoPE removes RoPE extension tricks, and Attention Residuals let layers retrieve earlier block representations instead of pushing everything through one residual stream.

K3's technical report reads like a catalog of ways to keep a 3T-class sparse model numerically well-behaved. suchenzang's scaling taxonomy bins the hard parts into sequence length, depth, and width.

The stability moves are concrete:

• KDA and Gated MLA mix local-ish and global attention in a 3:1 ratio, with qk-norm and channel-wise decay in suchenzang's KDA notes.
• Lower-bounded decay moves fragile inverse terms into log space and smooth truncation, according to suchenzang on finite precision.
• SiTU-GLU uses tanh * sigmoid to bound activations while approximating SwiGLU, stochasticchasm's SiTU note says.
• The final MoE fix stack is norm, bounded activations, and quantile balancing, in suchenzang's stability list.

That is the nerd payoff in this release: K3 is huge, but the report is mostly about not letting routing, activations, recurrent state, and low precision drive training into a ditch.

The report discloses model structure in detail, then gets quiet around training scale.

• stochasticchasm's compute note says training tokens and compute are not disclosed for K3 or MoonViT-V2.
• cedric_chee's comparison says K2.5 disclosed NVIDIA H800 clusters with 8 × 400 Gbps RoCE interconnects, while K3 is “very light” on hardware details.
• stochasticchasm's data note says pretraining data details are much thinner than in K2.

That opacity matters because Moonshot's 2.5x scaling-efficiency claim mixes architecture, data, and training improvements. Without tokens, FLOPs, and hardware, the split between those factors stays fuzzy.

Artificial Analysis puts K3 at 57 on its Intelligence Index, ahead of every other open-weight model it has tested and four points behind the leading closed models in ArtificialAnlys's comparison. Its model page also reports $3 per 1M input tokens, $15 per 1M output tokens, 32.6 output tokens per second, and unusually high verbosity on its eval run.

Agentic evals moved too:

• Agent Arena's leaderboard update ranks Kimi K3 Max first among open-weight models in Agent Arena with +9.75% net improvement, plus first-open rankings in Frontend Code and Text Arenas.
• cognition's Devin note says K3 is available in Devin Desktop and CLI, with 58.2% on FrontierCode 1.1 Extended and a 63.6% pass rate.
• FactoryAI's testing note says K3 performed on par with frontier labs in early coding-agent work, especially repo navigation, tool use, current-information lookup, and failure diagnosis.

The benchmark caveat is in Moonshot's own model card: K3 is often evaluated in Kimi Code or Claude Code harnesses, while OpenAI models use Codex on several coding benchmarks. Harness choice is part of the score now.

Serving K3 fast became a parallel launch. lmsysorg's SGLang post reported 423 tok/s on GSM8K with fused KDA decode kernels, DP attention, DSpark, P/D disaggregation, and KDA-aware prefix caching.

vLLM's day-zero thread is the more useful systems inventory:

• P/D disaggregation.
• TP8 prefill to DP16/EP16 decode over NIXL.
• MoE backends by topology.
• KV offloading for agent turns.
• Tool calling, reasoning, and structured output.
• DSpark speculative decoding, moving 118 tok/s to 370 tok/s on reasoning workloads, according to vLLM project's thread.

Modal claimed an even higher production number: modal's launch note says its custom DFlash speculator runs K3 faster without quality loss, while Modal's blog states 460 tokens per second on release day.

Pricing did not collapse on day zero. jaminball's pricing note says Baseten and Fireworks matched Moonshot's $3 input and $15 output per million tokens, despite expectations that open weights would immediately undercut the first-party API.

Open weights turned into immediate distribution across inference providers, agent products, and coding tools.

• Vercel AI Gateway added moonshotai/kimi-k3 and moonshotai/kimi-k3-fast from US-based providers with zero data retention, according to vercel_dev's gateway note.
• Factory added K3 to Droid at 50% off for two weeks in FactoryAI's launch note.
• Cognition added K3 to Devin Desktop and CLI in cognition's Devin note.
• Ollama Cloud exposed it for Claude Code through ollama launch claude --model kimi-k3:cloud, per ollama's cloud note.
• OpenRouter listed K3 with a growing set of third-party providers in OpenRouter's market note.
• Together AI launched K3 for long-running agent workflows in togethercompute's launch note.
• OpenCode Zen added K3 in opencode's post.
• Cursor added K3 with US-based inference and ZDR support in cursor_ai's launch note.
• Julius added K3 as its first open-weight model, citing large document sets, deep research, reports, dashboards, and PowerPoints in juliusai's launch note.

K3 is downloadable, but the first LocalLLaMA reaction called it an “F1 machine inside a show window.” the Reddit post says even 2 to 4 RTX 6000 Blackwell local workstations are not enough for native use.

iScienceLuvr's local note makes the same point: K3 is huge, so home use mostly means waiting for third-party APIs, optimizations, distillations, or smaller variants. teortaxesTex's local-user note puts it more bluntly: these models are built for datacenters.

The local ecosystem still moved. the Unsloth GGUF thread says Unsloth released GGUF files, and TheZachMueller's storage note frames the weight dump as a 40TB RAID moment.

AgentENV is the weirdest infrastructure detail in the release. Moonshot describes it as a distributed system for large-scale agent environments, with fast snapshot, resume, and fork support for parallel agent workflows.

The security boundary is the important new fact. rauchg's sandbox note says container-level isolation was not enough in Kimi's experiments because agents crashed the underlying machine through kernel panics, while Firecracker microVMs were safe.

stochasticchasm's sandbox note highlights pause and resume as the standout capability, with forking implying long-tail branching rollouts. That turns “agent environment” from a benchmark wrapper into persistent state infrastructure for million-token RL.