Kimi K3

Moonshot AI's flagship: a 2.8T-parameter MoE with 104B active, 1M context and native vision. Open weights shipped July 27, 2026.

Last reviewed September 6, 2026

At a glance

Overview

Kimi K3 is Moonshot AI’s flagship large language model, announced on July 16, 2026 (July 17 Beijing time) and live across Kimi.com, Kimi Work, Kimi Code and the API the same day. It displaced K2.6 immediately and completely: the Kimi API documentation calls K3 “Kimi’s most capable model to date,” the exact phrase it previously used for K2.6.

K3 is a 2.8-trillion-parameter Mixture-of-Experts model with 104 billion parameters active per token — the number that decides whether anyone can serve it. Architecturally it is not a scaled-up K2: it is built on Kimi Delta Attention (KDA), a hybrid linear attention mechanism, and Attention Residuals (AttnRes), with sparsity handled by a Stable LatentMoE framework selecting 16 of 896 routed experts. Moonshot claims roughly 2.5x the overall scaling efficiency of K2.

The weights shipped on July 27, 2026, on the date Moonshot named at launch, together with a technical report. That makes K3 the largest openly available model by a wide margin — with a licence and a hardware bill that both deserve reading before anyone plans around it.

Two other things matter for integration. The context window is 1,048,576 tokens, four times K2.6’s 256K, priced flat — there is no tiering by context length, which is how most vendors claw back the cost of a long window. And thinking is not optional: K3 always reasons, though reasoning_effort now offers three depths rather than the single max available on launch day.

Capabilities

On Moonshot’s own 35-benchmark table, K3 beats every listed competitor on seven rows. Those seven are the honest answer to “what is this model for”:

• Agentic web research: 91.2 on BrowseComp and 95.0 on DeepSearchQA — the clearest wins in the suite, both ahead of Claude Fable 5 and GPT 5.6 Sol.
• Long-horizon software engineering: 42.0 on SWE Marathon against 40.0 for Claude Opus 4.8 and 39.0 for GPT 5.6 Sol.
• Document parsing: 91.1 on OmniDocBench, the strongest vision result relative to the field.
• Binary-to-source reconstruction: 77.8 on Program Bench, though only 0.2 ahead of GPT 5.6 Sol and drawn from a different harness — see the benchmarks section.
• Computer-use automation: 30.8 on Automation Bench, on the 600-task public subset.
• 1M context, flat-priced: Long-repository and long-trajectory work without a context-length surcharge.
• Tunable reasoning depth: reasoning_effort accepts low, high and max — the one capability K3 gained after launch, when only max existed.

Elsewhere it is competitive rather than leading. GPQA-Diamond 93.5 ties GPT 5.5 and trails GPT 5.6 Sol’s 94.1; MMMU-Pro 81.6 and MathVision 94.3 both land third behind Fable 5 and Sol.

Technical Specifications

• API model ID: kimi-k3
• Total parameters: 2.8T
• Active parameters: 104B per token
• Experts: 896 routed, 16 selected per token, plus 2 shared (Stable LatentMoE)
• Layers: 93 — 1 dense, 69 Kimi Delta Attention (KDA), 24 Gated MLA
• Attention: KDA, a hybrid linear attention mechanism, plus Attention Residuals (AttnRes)
• Other stated components: Quantile Balancing, Per-Head Muon, Sigmoid Tanh Unit (SiTU)
• Numerics: MXFP4 weights with MXFP8 activations, quantization-aware trained
• Vision encoder: MoonViT-V2, 401M parameters; covers images and video natively
• Context window: 1,048,576 tokens
• Max completion tokens: defaults to 131,072; configurable up to 1,048,576
• Thinking: always on; reasoning_effort accepts low, high, max (default max)
• Fixed sampling parameters: temperature=1.0, top_p=0.95, n=1, presence_penalty=0, frequency_penalty=0 are fixed — omit them
• Weights: released July 27, 2026 under the Kimi K3 License
• Checkpoint size: ~1.56 TB across 96 safetensors shards
• API compatibility: OpenAI-compatible endpoints with prompt caching

Moonshot does not publish a knowledge cutoff for K3, so this page does not state one.

What actually shipped on July 27

Moonshot published moonshotai/Kimi-K3 on Hugging Face on July 27, 2026, with a 47-page technical report covering architecture, pre-training, post-training, infrastructure and evaluations. What is in the repository is a native MXFP4 checkpoint — roughly 1.56 TB across 96 shards. There is no BF16 or FP16 release and there is not meant to be one: quantization-aware training was applied from supervised fine-tuning onward, so MXFP4 is the model’s trained precision, not a lossy post-hoc conversion. At FP16 a 2.8T model would be about 5.6 TB, which is the practical reason the release looks the way it does.

The licence is the Kimi K3 License, and it is not the Modified MIT the K2 line shipped under. Hugging Face tags it license:other. Internal deployment, research and ordinary product use are free and require no permission. Two revenue-triggered conditions apply on top:

• A Model-as-a-Service operator whose group revenue exceeds $20 million over any consecutive 12 months must sign a separate agreement with Moonshot before commercially serving K3 inference or fine-tuning.
• Any commercial product with more than 100 million monthly active users or more than $20 million in monthly revenue must display “Kimi K3” prominently in its interface.

Both conditions are waived for purely internal use and for access through Moonshot’s own products or certified inference partners. This is open weights, not open source, and Moonshot has been consistent about saying so.

Serving it is the harder gate. Moonshot recommends 64 or more accelerators; the smallest configuration validated by the vLLM project is 16 GPUs (DP16+EP16). Below datacenter-scale capacity, self-hosting K3 is not a plan.

The K2 line

kimi-k2.6 and kimi-k2.7-code remain callable and are still the right answer for some work.

K2.6 is a 1T MoE with 32B active per token, 384 experts with 8 selected, 61 layers and a 400M-parameter MoonViT encoder. K2.7-Code is “a coding-focused agentic model built upon Kimi K2.6” — not a general upgrade, and there is no general-purpose kimi-k2.7. It scores 62.0 on Kimi Code Bench v2 against K2.6’s 50.9.

Retired models

Five kimi-k2* preview models — kimi-k2-0905-preview, kimi-k2-0711-preview, kimi-k2-turbo-preview, kimi-k2-thinking and kimi-k2-thinking-turbo — were discontinued on May 25, 2026.

K3’s launch triggered the next round, and it has now completed. kimi-k2.5 and the entire moonshot-v1 series — including moonshot-v1-auto and the -vision-preview variants — were officially retired on August 31, 2026. Calls to those IDs return a 404 model-not-found; Moonshot’s migration target is kimi-k3. The K2.5 retirement has a downstream footnote worth knowing: Cursor’s Composer 2.5 is post-trained from Moonshot’s open K2.5 checkpoint, which remains downloadable and unaffected — it is the hosted kimi-k2.5 endpoint that is gone, not the open weights it was built from.

Use Cases

• Deep research agents: K3’s most defensible strength. BrowseComp 91.2 and DeepSearchQA 95.0 are wins over both Fable 5 and Sol, and multi-hop research is where a 1M window and mandatory reasoning both pay for themselves.
• Long-running coding sessions: SWE Marathon 42.0 leads the field — with the caveat that the same long-horizon training produces the proactiveness problem below.
• Million-token repository analysis: Flat pricing across the full 1M window makes whole-monorepo passes economically predictable in a way tiered vendors do not.
• Document and video understanding: OmniDocBench 91.1 tops the table, and native video vision removes a separate frame-sampling pipeline.
• Sovereign and air-gapped deployment: The real payoff of the July 27 release. If you already run a 64-accelerator cluster and need frontier-adjacent capability inside your own network, K3 is now the strongest option that exists — subject to the licence’s revenue gates.
• What to route elsewhere: High-volume, latency-sensitive or shallow work. Even at reasoning_effort: low, K3 reasons before answering, and $15.00 per 1M output makes it a poor fit for classification, extraction or chat, where kimi-k2.7-code or a smaller model is four times cheaper. For pure reasoning or vision benchmarks, Fable 5 and Sol simply score higher.

Performance / Benchmarks

All figures below are vendor-run, published by Moonshot in the July 16, 2026 announcement and reproduced unchanged in the July 27 technical report — the report added disclosure, not new numbers. K3 is measured at reasoning_effort: max, temperature=1.0, top-p=1.0. A representative selection; the full table is in the announcement and the technical report.

The scoreboard, counted honestly: K3 beats every listed competitor on 7 of 35 rows and loses to at least one on 27. Claude Fable 5 takes 20 of those 27, GPT 5.6 Sol the other 7. This is why Moonshot’s own framing matters — K3 shows “frontier-level performance across our evaluation suite, consistently outperforming other tested models,” but “still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol.” A lab conceding second place in its own launch post is worth more than any table.

Three caveats survive the technical report:

The harnesses differ. Moonshot states that “depending on the benchmark, each model is evaluated under one of three agentic harnesses — KimiCode, Claude Code, or Codex.” K3 mostly runs on KimiCode while competitors are scored under Claude Code or Codex. On Terminal Bench 2.1, K3’s 88.3 is a KimiCode run while every competitor gets “the best score across harnesses” — not like-for-like in either direction.

Many competitor numbers are imported, not measured. Per-row footnotes cite vals.ai, frontierswe.com, Artificial Analysis and the vendors’ own posts; Kimi Code Bench 2.0, DECK-Bench and PerceptionBench are in-house benchmarks, and GPQA-Diamond and HLE-Full carry no sourcing footnote at all.

Some margins are noise. Program Bench (77.8 vs 77.6) and SpreadsheetBench 2 (34.8 vs 34.7) are wins by a rounding error across incompatible harnesses.

No comparison with K2.6 is possible here — K3’s suite shares no rows with the K2.6 model card (SWE-Bench Verified 80.2, Terminal-Bench 2.0 66.7, BrowseComp 83.2). Now that the weights are public, third-party evaluation is possible for the first time; run your own.

Limitations

• Open weights, gated licence: The Kimi K3 License is not MIT. If you resell inference at scale, or ship to 100M+ monthly users, read the two revenue clauses in Technical Specifications before you deploy.
• Self-hosting needs a cluster: ~1.56 TB of MXFP4 weights before KV cache, a vendor recommendation of 64+ accelerators, and a 16-GPU floor validated by vLLM. “Openly available” and “runnable by you” are different claims.
• MXFP4 is the only precision offered: There is no higher-precision checkpoint to fall back on, and no vendor-published ablation of what the quantization-aware training costs. Fine-tuning workflows that assume BF16 weights need rethinking.
• Thinking cannot be turned off: reasoning_effort now spans low, high and max, which is a real improvement over launch day, but there is no non-reasoning path through K3 and reasoning tokens bill as output.
• Roughly 4x K2.6’s output price: $15.00 against $4.00 per 1M output, compounded by mandatory reasoning tokens. The real per-task delta is larger than the rate card suggests.
• Excessive proactiveness: Vendor-acknowledged. On minor issues or ambiguous intent, K3 “may make unexpected decisions on the user’s behalf” — a consequence of long-horizon training that matters in agent loops with side effects. Moonshot’s advice is to constrain it via the system prompt or an AGENTS.md.
• Sensitivity to thinking history: Vendor-acknowledged, and the sharpest operational edge. The docs require you to “return the complete assistant message unchanged in multi-turn conversations and tool calls.” If your harness drops historical thinking content, or you switch a session from another model to K3 mid-flight, generation quality can become highly unstable.
• A user-experience gap, per Moonshot: The vendor’s own third stated limitation is “a noticeable gap in user experience compared with Claude Fable 5 and GPT 5.6 Sol.”
• Sampling parameters are locked: temperature, top_p, n and both penalties are fixed. Prompt engineering is your only lever.
• No published knowledge cutoff: Ground time-sensitive queries with retrieval or tool use.
• Regional availability: Hosted API access may be restricted in some jurisdictions.

Pricing & Access

From the Kimi platform pricing pages, per 1M tokens in USD, checked September 6, 2026. Launch pricing is unchanged.

K3 pricing is flat pay-as-you-go with no tiering by context length — the millionth token costs what the first one does. Cache hits run one-tenth of a cache miss, a wider spread than on K2.6, so prompt caching is the single highest-leverage optimization for any agent replaying a long system prompt or repository context.

If you are still on a retired model ID: kimi-k2.5 and the whole moonshot-v1 series (including moonshot-v1-auto and the -vision-preview variants) were officially retired on August 31, 2026 and now return a 404. Moonshot’s stated migration target is kimi-k3; if the price step is too large, kimi-k2.6 is the closer like-for-like on cost and is still supported. The older kimi-k2* previews went on May 25, 2026 and are equally gone.

Access options:
– Kimi API platform — OpenAI-compatible endpoints, kimi-k3
– Kimi Code — the CLI and coding agent, where the model ID is plain k3. Access is plan-gated: unavailable on Andante, a 256K window on Moderato, and the full 1M on Allegretto and above. Its other two IDs, kimi-for-coding and kimi-for-coding-highspeed, are both K2.7 Code — the same model, with the high-speed variant running ~6x faster at 3x quota
– Kimi.com and Kimi Work — consumer and workspace applications
– Hugging Face — the K3 MXFP4 checkpoint, plus K2.6 and K2.7-Code weights
– Self-hosting — vLLM and SGLang have community serving paths for K3 and the K2 line

Ecosystem & Tools

• Kimi-K3 on Hugging Face — the released weights, the model card, and the Kimi K3 License text
• Kimi-K3 on GitHub — the technical report and the architecture disclosure
• Kimi K3 quickstart — the authoritative source on reasoning_effort, fixed sampling parameters, and completion limits
• Kimi API model list — every callable model ID, context length, and retirement notice
• Kimi-K2.6 on Hugging Face — the previous flagship’s weights and full benchmark table
• Cursor Composer 2.5 — post-trained on Moonshot’s open Kimi K2.5 checkpoint, the most visible downstream use of the open weights
• vLLM and SGLang — community serving paths; vLLM has validated a 16-GPU K3 configuration

Community & Resources

• Kimi K3 announcement — July 16, 2026, with the full benchmark table and stated limitations
• Kimi K3 technical report — published July 27, 2026 alongside the weights
• Moonshot AI — release timeline
• Kimi platform documentation
• Compare with DeepSeek V4, GLM-5.3, Qwen3.8-Max, and Ling-2.6-1T