Kimi K3 on Run BiOS

Kimi K3 is an open-weight, natively multimodal agentic model developed by Moonshot AI, which the company describes as its most capable model to date. It is a 2.8-trillion-parameter Mixture-of-Experts (MoE) model built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with native vision capabilities and a 1-million-token context window. Moonshot AI presents it as the world's first open 3T-class model, designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning.

Moonshot AI reports that, while overall performance still trails the most powerful proprietary models the company evaluated (Claude Fable 5 and GPT-5.6 Sol), Kimi K3 achieves frontier-level performance across its evaluation suite and consistently outperforms the other open and proprietary models tested.

Key properties reported by Moonshot AI:

• New architecture. Kimi K3 is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), and scales up MoE sparsity with a Stable LatentMoE framework that activates 16 out of 896 experts — yielding, by Moonshot AI's scaling-law studies, an approximate 2.5× improvement in overall scaling efficiency over Kimi K2.
• Long-horizon coding. Operating with minimal human oversight, Kimi K3 sustains long engineering sessions, navigates massive repositories, and orchestrates terminal tools — from GPU kernel optimization and compiler development to vision-in-the-loop game development, CAD, and even chip design. In one published case study, an early version of Kimi K3 handled the majority of the Kimi team's own kernel-optimization work; in another, the model designed and verified a chip in a single 48-hour autonomous run using open-source EDA tools on the Nangate 45nm library.
• Agentic knowledge work. Kimi K3 advances end-to-end knowledge work, producing deep research with interactive visualizations, widgets and dashboards, and motion design and video editing, powered by its native multimodal architecture.
• Native multimodality and long context. Kimi K3 understands text, images, and video within the same model and supports a 1-million-token context window.
• Open frontier weights. Moonshot AI releases the full Kimi K3 model weights under the Kimi K3 License, making frontier intelligence openly available for research, deployment, and further innovation.

Architecture

All architecture figures are from Moonshot AI's official model card.

Property	Value
Architecture	Mixture-of-Experts (MoE)
Total Parameters	2.8T
Activated Parameters	104B
Number of Layers	93
Number of Dense Layers	1
Attention-Layer Composition	69 KDA + 24 Gated MLA
Attention Hidden Dimension	7168
Number of Attention Heads	96
Latent MoE Dimension	3584
MoE Hidden Dimension (per Expert)	3072
Number of Experts	896
Selected Experts per Token	16
Number of Shared Experts	2
Vocabulary Size	160K
Context Length	1048576
Attention Mechanism	KDA & Gated MLA
Activation Function	SiTU-GLU
Vision Encoder	MoonViT-V2
Parameters of Vision Encoder	401M
Quantization	MXFP4 weights / MXFP8 activations (quantization-aware training)
Modality	Text, Image

Per Moonshot AI's technical report, each block combines three KDA layers with one Gated MLA layer; Attention Residuals let each module selectively retrieve representations from the embedding and preceding blocks rather than accumulating them uniformly. Stable LatentMoE performs sparse channel mixing, effectively activating 16 of 896 routed experts per token. Routing stability at this sparsity comes from Quantile Balancing (which derives expert allocation directly from router-score quantiles) and Per-Head Muon (which optimizes attention heads independently). MoonViT-V2 encodes images and videos, and a lightweight projector maps visual features into the shared embedding space. Kimi K3 applies quantization-aware training from the SFT stage onward, using MXFP4 weights with MXFP8 activations for broad hardware compatibility.

Training

Reported by Moonshot AI's technical report:

• Pre-training data. A curated corpus spanning four primary text domains — Web Text, Code, Mathematics, and Knowledge — plus a large-scale vision corpus covering captions, interleaved image–text documents, OCR, perception, video, and visual coding data. Each text domain is filtered with rule-based heuristics, classifier-based quality scoring, and deduplication; knowledge and mathematics corpora are rephrased with style- and perspective-diverse prompting and fidelity verification against source documents. The vision corpus combines open collections with in-house filtering, synthesis, and deduplication pipelines, and substantially scales up programmatic multimodal data (code snippets coupled with rendered visuals across SVG, 3D assets, webpages, games, and CAD schematics).
• Training schedule. Scaling-law studies favored cosine decay over Warmup Stable Decay. Pre-training begins at a context length of 8K tokens, later extends to 64K, and the window grows from 256K to 1M tokens during the cooldown phase — a four-stage curriculum that concentrates costly long-sequence computation within a small fraction of the overall training budget.
• Post-training. Reinforcement learning across general, agentic, and coding domains at multiple reasoning-effort levels. Training environments include verifiable search and professional knowledge work, software engineering and kernel optimization, multimodal reasoning with vision-in-the-loop tool use, persistent assistant workflows, web development, and autonomous execution tasks — training a general loop of reasoning, acting, observing, verifying, and adapting, often over hundreds or thousands of tool calls and millions of accumulated context tokens. Domain- and effort-specialized policies are consolidated into a single unified model through multi-teacher on-policy distillation.
• Multi-token prediction. Kimi K3 is pre-trained with a multi-token-prediction (MTP) layer that mirrors the structure of a backbone block; Moonshot AI fine-tunes it into an EAGLE-3-style draft model for accelerated decoding.

Benchmarks

All results below are from Moonshot AI's official model card. Per the card: all Kimi K3 results are obtained with reasoning effort set to "max" and temperature = 1.0. For single-step tasks, such as GPQA Diamond, HLE-Full, and vision benchmarks without tools, top-p = 0.95; for agentic tasks, top-p = 1.0. For HLE-Full, MMMU-Pro, CharXiv (RQ), MathVision, and ZeroBench, each cell reports the scores without and with tool augmentation (general tools for HLE-Full, Python for the vision benchmarks), in that order. "—" means the score was not reported.

Reasoning & Knowledge

Benchmark	Kimi K3 (max)	Claude Fable 5 (max, w/ fallback)	GPT-5.6 Sol (max)	Claude Opus 4.8 (max)	GPT-5.5 (xhigh)	GLM-5.2 (max)
GPQA Diamond	93.5	92.6	94.1	91.0	93.5	91.2
CritPt	23.4	28.6	32.3	20.9	27.1	20.9
AA-LCR	74.7	70.0	73.7	67.7	74.3	71.3
HLE-Full	43.5 / 56.0	53.3 / 63.0	44.5 / 58.0	49.8 / 57.9	41.4 / 52.2	—

Coding

Benchmark	Kimi K3 (max)	Claude Fable 5 (max, w/ fallback)	GPT-5.6 Sol (max)	Claude Opus 4.8 (max)	GPT-5.5 (xhigh)	GLM-5.2 (max)
DeepSWE	67.5	70.0	73.0	59.0	67.0	46.2
ProgramBench	77.8	76.8	77.6	71.9	70.8	63.7
Terminal-Bench 2.1	88.3	88.0	88.8	84.6	83.4	82.7
FrontierSWE	81.2	86.6	71.3	66.7	64.9	67.3
SWE-Marathon	42.0	35.0	39.0	40.0	14.0	13.0
PostTrainBench	36.6	41.4	34.6	34.1	28.4	34.3
MLS-Bench-Lite	48.3	49.9	46.2	42.8	35.5	40.4
SciCode	58.7	60.2	56.1	53.5	56.1	50.5
Kimi Code Bench 2.0	72.9	76.9	64.8	71.7	69.0	64.2

Agentic

Benchmark	Kimi K3 (max)	Claude Fable 5 (max, w/ fallback)	GPT-5.6 Sol (max)	Claude Opus 4.8 (max)	GPT-5.5 (xhigh)	GLM-5.2 (max)
BrowseComp	91.2	88.0	90.4	84.3	84.4	—
DeepSearchQA (F1)	95.0	94.2	—	93.1	—	—
ResearchRubrics	76.2	—	73.8	73.5	64.0	71.1
GDPval-AA v2 (Elo)	1686	1747	1736	1593	1491	1510
Toolathlon-Verified	76.5	77.9	74.9	76.2	73.5	59.9
MCPMark-Verified	94.5	87.4	92.9	76.4	92.9	—
MCP-Atlas	84.2	84.7	83.6	83.6	82.8	82.6
AutomationBench	30.8	29.1	29.7	27.2	22.7	12.9
JobBench	54.3	57.4	45.4	48.4	38.3	43.4
AA-Briefcase (Elo)	1548	1583	1495	1354	1158	1260
Agents' Last Exam	28.3	25.7 †	29.6	27.0	26.6	20.4
APEX-Agents	41.0	43.3	39.9	39.4	38.5	35.6
OfficeQA Pro	63.3	69.9	63.2	63.9	60.9	41.4
SpreadsheetBench 2	34.8	34.7	32.4	31.6	29.1	28.1
OSWorld-Verified	84.8	85.0	83.0	83.4	79.0	—
OSWorld 2.0	58.3	66.1	62.6	55.7	49.5	—
SaaS-Bench	60.1	—	61.4	56.1	43.8	—
τ³-Banking	33.4	26.8	33.0	27.6	31.3	26.8
Harvey Lab-AA	94.6	93.6	87.2	91.1	86.3	91.0
CorpFin v2	71.6	71.8	64.4	66.7	68.4	66.1
Finance Agent v2	54.4	56.3	53.8	53.9	51.8	49.7
Legal Research Bench	44.2	49.5	48.1	43.8	40.4	31.3

Vision

Benchmark	Kimi K3 (max)	Claude Fable 5 (max, w/ fallback)	GPT-5.6 Sol (max)	Claude Opus 4.8 (max)	GPT-5.5 (xhigh)	GLM-5.2 (max)
WorldVQA ForceAnswer	51.0	56.7	41.8	39.1	38.5	—
OmniDocBench	91.1	89.8	85.8	87.9	89.4	—
PerceptionBench	58.5	57.2	59.7	47.2	55.8	—
Video-MME (w. sub)	90.0	—	89.5	86.0	89.3	—
MMVU	82.1	—	81.2	79.2	81.7	—
BabyVision w/ python	85.7	90.5	88.9	81.2	83.6	—
MMMU-Pro	81.6 / 83.4	81.2 / 86.5	83.0 / 84.6	78.9 / 82.7	81.2 / 83.2	—
CharXiv (RQ)	84.8 / 91.3	88.9 / 93.5	84.6 / 89.1	80.5 / 89.9	84.1 / 89.0	—
MathVision	94.3 / 97.8	94.8 / 98.6	95.8 / 97.8	86.7 / 97.1	92.2 / 96.8	—
ZeroBench (pass@5)	23.0 / 41.0	23.0 / 46.0	17.0 / 35.0	17.0 / 34.0	22.0 / 41.0	—

† Per Moonshot AI's footnote, the Claude Fable 5 entry on Agents' Last Exam runs at xhigh effort with 40% of tasks annotated as downgraded.

Benchmark methodology notes (as disclosed by Moonshot AI)

• CritPt and AA-LCR. Scores are cited from Artificial Analysis as of July 23, 2026.
• DeepSWE. Kimi K3 is evaluated with the Kimi Code harness (it attains 67.3 with the mini-SWE-agent harness on the official DeepSWE leaderboard). The GLM-5.2 score is from the GLM-5.2 release blog; remaining scores are from the official DeepSWE leaderboard. Moonshot AI reports the DeepSWE v1.1 tasks.
• Terminal-Bench 2.1. Kimi K3 is evaluated with the Kimi Code harness. For all other models, Moonshot AI reports the best score across harnesses: GLM-5.2 with Claude Code (from the GLM-5.2 release blog); Claude Opus 4.8 and Claude Fable 5 with Terminus 2 (from Artificial Analysis); GPT-5.5 and GPT-5.6 Sol with Codex (from OpenAI).
• ProgramBench. Kimi K3 is evaluated with the Kimi Code harness. The GLM-5.2 score is from the GLM-5.2 release blog; all other scores are from Vals AI.
• SWE-Marathon. Kimi K3, Claude Opus 4.8, and Claude Fable 5 are evaluated with the Claude Code harness; GPT-5.6 Sol with the Codex harness; the GLM-5.2 score is from the GLM-5.2 release blog. Moonshot AI's evaluation is based on an H20-calibrated branch of the official tasks as of July 9, 2026, prior to the final v1.1 release: Docker images, performance gates, and reference oracles for the GPU tasks were recalibrated for H20, while correctness and anti-cheat validators remain unchanged. Claude Fable 5 hit fallbacks on 35% of the tasks in this evaluation, which may have negatively impacted its measured performance.
• FrontierSWE. Kimi K3 is evaluated with the Kimi Code harness and GPT-5.6 Sol with the Codex harness; all other results are from FrontierSWE. Dominance scores are recomputed from raw scores using the official evaluation script and are current as of July 16, 2026.
• PostTrainBench. Scores for GLM-5.2, GPT-5.5, and Claude Opus 4.8 are adopted from the official PostTrainBench results. Kimi K3, Claude Fable 5, and GPT-5.6 Sol are evaluated with the official Harbor implementation at maximum reasoning effort, averaged over three runs on H20 GPUs (instead of H100 in the official setting) — Kimi K3 and Claude Fable 5 with the Claude Code harness, and GPT-5.6 Sol with the Codex harness.
• MLS-Bench-Lite. Kimi K3 is evaluated with the Kimi Code harness; GLM-5.2 and the Claude models with the Claude Code harness; GPT-5.5 and GPT-5.6 Sol with the Codex harness.
• SciCode. Scores are cited from Artificial Analysis as of July 23, 2026.
• Kimi Code Bench 2.0 (in-house). Kimi K3 is evaluated with the Kimi Code harness (it attains 73.7 with the Claude Code harness); GLM-5.2, Claude Opus 4.8, and Claude Fable 5 with the Claude Code harness; GPT-5.5 and GPT-5.6 Sol with the Codex harness. All models are evaluated at maximum reasoning effort, except GPT-5.5, which uses the "xhigh" setting. As the benchmark includes cybersecurity and safety-related tasks, Moonshot AI also discloses the fraction of refused or fallback tasks: Claude Fable 5 hit 13 fallbacks and 1 refusal out of 80 tasks; 10 refusals out of 80 tasks entered GPT-5.6 Sol's cyber guard; GPT-5.5 had 3 refusals out of 80 tasks.
• OfficeQA Pro. Each test case provides the agent with the entire PDF corpus, with all PDFs rendered as images and no machine-readable text available.
• OfficeQA Pro and SpreadsheetBench 2. Kimi K3, GLM-5.2, Claude Opus 4.8, and Claude Fable 5 are evaluated with the Claude Code harness; GPT-5.5 and GPT-5.6 Sol with the Codex harness.
• MCP-Atlas. All models are evaluated on the 500-task public subset with a 100-turn limit, using Gemini 3.1 Pro as the judge.
• AutomationBench. All models are evaluated on the 600-task public subset.
• BrowseComp. Moonshot AI adopts a context-compaction strategy triggered at 300K tokens. When evaluated with the full 1M-token context window and no context management, Kimi K3 achieves a score of 90.4. The results of Claude Fable 5, Claude Opus 4.8, GPT-5.6 Sol, and GPT-5.5 are cited from Anthropic and OpenAI.
• GDPval-AA v2, AA-Briefcase, τ³-Banking, Harvey Lab-AA, and APEX-Agents. Scores are cited from Artificial Analysis and the APEX-Agents leaderboard as of July 23, 2026. For Harvey Lab-AA, Moonshot AI reports the criterion pass rate.
• CorpFin v2, Finance Agent v2, and Legal Research Bench. Scores are cited from Vals AI.
• Agents' Last Exam. Scores are cited from the official leaderboard as of July 23, 2026; Moonshot AI reports the leaderboard's primary pass-rate metric. On the leaderboard, each model is paired with a specific harness: Kimi K3 with Kimi Code; GPT-5.6 Sol and GPT-5.5 with Codex; Claude Fable 5, Claude Opus 4.8, and GLM-5.2 with Claude Code.
• Multimodal benchmarks. Except for ZeroBench, which follows the official setting and is run five times, all multimodal scores are averaged over three runs. MMMU-Pro is evaluated following the official protocol, preserving the original input order and prepending images to the text input. PerceptionBench is an in-house benchmark that focuses on atomic visual-perception capabilities.

Thinking and Tool Use on Run BiOS

• Reasoning is always on. Per Moonshot AI's documentation, Kimi K3 always has thinking enabled and returns its reasoning content alongside each response; there is no non-thinking mode. Thinking effort is configurable with low, high, and max levels (max is the vendor default and the setting used for all of the benchmark results above).
• Preserved thinking history. Kimi K3 was trained in the preserved-thinking-history mode: for multi-turn conversations and tool calls, the complete assistant message — including reasoning content and tool calls, not just the final text — is passed back on each turn.
• Function calling. Kimi K3 supports structured tool use, including interleaved thinking and multi-step tool calls, for agentic workflows.
• Vision input. The model accepts image input natively (Moonshot AI's model card lists the modality as text and image; its technical report describes the MoonViT-V2 encoder as handling images and video).
• Context window. 1,048,576 tokens (1M), matching the model card's published context length of 1048576.
• Prompt caching. Repeated prefixes are served from cache at a reduced rate.

Safety, Security Evaluation, and Limitations

Disclosed by Moonshot AI across its model card, tech blog, and technical report:

• Cyber-security evaluation. Moonshot AI evaluated Kimi K3's cybersecurity capability along a two-tier progression of increasing operational risk: vulnerability discovery with proof-of-concept development (Tier 1, primarily associated with defensive security research) and end-to-end exploit development (Tier 2, the tier most directly relevant to misuse risk). At Tier 1, across dozens of widely deployed systems (operating-system kernels, databases, AI services, web frameworks, blockchain, and VPN software), the model identified hundreds of candidate vulnerabilities; of the findings that underwent human review, approximately 70% were confirmed genuine, including 16 previously unknown vulnerabilities across six projects — among them a remotely triggerable heap out-of-bounds write in the Linux kernel (confirmed as a remote denial-of-service primitive) and a Dirty-COW-class vulnerability in the RDMA subsystem (confirmed as a deterministic local privilege-escalation primitive). At Tier 2, on an in-house suite of 36 expert-verified tasks (16 user-space exploitation tasks covering real CVEs in widely deployed software such as PostgreSQL, XWiki, and the Apache HTTP Server; 20 Linux-kernel exploitation tasks with progressively enabled mitigations), Kimi K3 solved 14 of 36 tasks (38.9%) versus 8 of 36 (22.2%) for GLM-5.2, with 10 of its 14 successes in the user-space track. Moonshot AI concludes that cyber capability is strongest at Tier 1 and user-space exploitation, while a clear gap to human experts remains on hardened targets. An independent joint assessment by the UK AI Security Institute and NIST's Center for AI Standards and Innovation (CAISI) reaches conclusions consistent with Moonshot AI's: Kimi K3 outperforms GLM-5.2 on exploit development (32% vs. 24% on ExploitBench; 17 vs. 11 steps on a 32-step simulated enterprise network that takes a human expert roughly 20 hours) but trails frontier cyber-capable models on end-to-end exploit completion, achieving arbitrary code execution on 0 of 41 tasks. Moonshot AI regards its evaluation as a lower bound on capability, conditioned on the current model version and evaluation coverage, and plans to revisit the results at each major model update.
• Reward-hacking safeguards. In kernel-optimization RL training, Moonshot AI developed a hacking-detection system that penalizes reward-hacking strategies such as CUDA graph replay, input caching, and precision reduction, and continuously extends it as new hacking strategies are observed.
• Benchmark-integrity disclosure. Because Kimi Code Bench 2.0 includes cybersecurity and safety-related tasks, Moonshot AI discloses refusal and fallback counts for the models evaluated (see the benchmark methodology notes above).

Limitations stated by Moonshot AI

1. Sensitivity to thinking history. Kimi K3 was trained in the preserved-thinking-history mode. If the agent harness fails to pass back all historical thinking content as required, or if an ongoing session with another model is switched over to Kimi K3, generation quality may become highly unstable. Moonshot AI recommends using a harness with verified compatibility and avoiding switching to Kimi K3 in the middle of a session.
2. Excessive proactiveness. Kimi K3's training places particular emphasis on long-horizon, challenging tasks. As a result, when it encounters minor issues or ambiguous user intent during task execution, it may make unexpected decisions on the user's behalf. Moonshot AI advises that applications requiring the agent to operate within well-defined boundaries impose more explicit behavioral constraints in the system prompt or an agent-instructions file.
3. Gap to frontier proprietary models. Despite being a highly competitive model overall, Kimi K3 nonetheless exhibits a noticeable gap in user experience compared with Claude Fable 5 and GPT-5.6 Sol, per Moonshot AI.

Availability

Kimi K3 is served on Run BiOS under the model id kimi-k3, with a 1,048,576-token context window, always-on reasoning with low/high/max effort levels, vision input, function calling, and prompt caching.