Kimi K3

Overview

Kimi K3 is Moonshot AI's flagship large language model, announced on July 16, 2026 (July 17 Beijing time). It displaced K2.6 immediately and completely: the Kimi API documentation now calls K3 "Kimi's most capable model to date," the exact phrase it previously used for K2.6, and points general work at kimi-k3.

The scale claim is the headline. K3 is a 2.8-trillion-parameter Mixture-of-Experts model, which Moonshot markets as "the world's first open 3T-class model" — 2.8T is the real number, "3T-class" the rounding. That phrasing is doing some work anyway, because the weights are not out yet. 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 that "effectively activat[es] 16 out of 896 experts." Moonshot claims the combination delivers "roughly 2.5x the overall scaling efficiency of K2."

Two changes matter more than the parameter count for anyone integrating it. The context window is now 1M tokens, four times K2.6's 256K, and it is 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 no longer optional: K3 always reasons, and reasoning_effort currently accepts only max.

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.

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
• Experts: 896, "effectively activating 16 out of 896" per token (Stable LatentMoE)
• Attention: Kimi Delta Attention (KDA), a hybrid linear attention mechanism, plus Attention Residuals (AttnRes) and Gated MLA
• Other stated components: Quantile Balancing, Per-Head Muon, Sigmoid Tanh Unit (SiTU)
• Numerics: MXFP4 weights with MXFP8 activations
• Context window: 1M tokens (1,048,576)
• Max completion tokens: defaults to 131,072; configurable up to 1,048,576
• Vision: native, covering images and video
• Thinking: always on; reasoning_effort supports only max today, more levels "coming soon"
• Fixed sampling parameters: temperature=1.0, top_p=0.95, n=1, presence_penalty=0, frequency_penalty=0 are fixed — omit them
• License: not stated; weights announced for July 27, 2026
• API compatibility: OpenAI-compatible endpoints with prompt caching

Moonshot does not publish a knowledge cutoff for K3, so this page does not state one. A full technical report is promised alongside the weights.

The open-weights claim

Moonshot calls K3 "the world's first open 3T-class model" and states that "the full model weights will be released by July 27, 2026." As of July 17, 2026 there is no Kimi-K3 repository on Hugging Face and no license has been named — K2.6 and K2.7-Code ship under a Modified MIT License, but nothing says K3 will.

So K3 is currently a proprietary API model with an open-weights promise attached. The promise has a track record behind it — the K2 line is genuinely open, and Cursor's Composer 2.5 is post-trained on the open Kimi K2.5 checkpoint — but a date is a plan, not a fact. Do not architect around self-hosted K3 until the repository exists.

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 vision 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 and reduces "thinking-token usage by approximately 30% compared with Kimi K2.6," which is exactly the property K3 removes by making thinking mandatory.

Deprecated models

Five kimi-k2* models — kimi-k2-0905-preview, kimi-k2-0711-preview, kimi-k2-turbo-preview, kimi-k2-thinking, and kimi-k2-thinking-turbo — were "officially discontinued on May 25, 2026 and are no longer maintained or supported." kimi-latest went on January 28, 2026 and kimi-thinking-preview on November 11, 2025.

K3's launch triggered the next round: "Following the Kimi K3 launch, kimi-k2.5 and the moonshot-v1 series are no longer available to newly registered users (full platform sunset on August 31)." Two distinct events, and the docs state no year for that August 31 sunset — existing users keep access until then, new ones already have none.

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, and long-horizon work is what Moonshot trained for — with the caveat that the same 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 vision covering video rather than stills removes a separate frame-sampling pipeline.
• Self-hosting, from July 27: If the weights land as promised, K3 becomes the largest openly available model by a wide margin — the reason to care even where a proprietary model scores higher.
• What to route elsewhere: High-volume, latency-sensitive, or shallow work. Mandatory max thinking at $15.00 per 1M output makes K3 a poor fit for classification, extraction, or chat, where K2.7-Code or a smaller model is four times cheaper and does not reason before answering. For pure reasoning or vision benchmarks, Fable 5 and Sol simply score higher.

Performance / Benchmarks

Moonshot's announcement publishes 35 benchmarks across six models. K3 is measured at reasoning_effort: max, temperature=1.0, top-p=1.0. A representative selection — full table in the announcement:

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.

To Moonshot's credit, that table is a fuller disclosure than most labs publish — which is also what makes its weaknesses visible. Three caveats:

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 Anthropic's and OpenAI's own posts; GDPval-AA v2 and AA-Briefcase come wholly from Artificial Analysis. Kimi Code Bench 2.0, DECK-Bench and PerceptionBench are in-house benchmarks. Toolathlon, Job Bench, APEX-Agents, 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, and twelve cells carry an asterisk the page never defines.

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). Validate on your own evaluation set.

Limitations

• The weights are not out: "Open 3T-class model" describes a promise dated July 27, 2026, not something you can download today. No license has been named.
• Thinking cannot be turned off: Every call reasons at max effort. There is no cheap path through K3 for simple tasks, and no way to trade accuracy for latency.
• 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: Also vendor-acknowledged, and the sharpest operational edge. K3 was trained in preserved-thinking-history mode; if your harness fails to pass back historical thinking content, or you switch a session from another model to K3 mid-flight, "generation quality may become highly unstable." Moonshot recommends Kimi Code and no mid-session switching.
• A user-experience gap, per Moonshot: The third stated limitation is that "despite being a highly competitive model overall, K3 nonetheless exhibits 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.
• Self-hosting will be brutal: A 2.8T MoE, once released, will need a substantially larger cluster than the 1T K2.6 that already required a serious multi-GPU rig.
• Regional availability: Hosted API access may be restricted in some jurisdictions.

Pricing & Access

From the Kimi platform pricing pages, per 1M tokens in USD.

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.

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 — K2.6 and K2.7-Code weights today; K3 weights promised by July 27, 2026

Ecosystem & Tools

• 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 deprecation notice
• Kimi-K2.6 on Hugging Face — weights and the full benchmark table
• Kimi-K2.7-Code on Hugging Face — coding-specialized weights
• 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 for the open K2 weights

Community & Resources

• Kimi K3 announcement — July 16, 2026, with the full benchmark table and stated limitations
• Moonshot AI — release timeline
• Kimi platform documentation
• Kimi K2.6 model card
• Compare with DeepSeek V4, GLM-5.2, Qwen3.7-Max, and Ling-2.6-1T