Kimi K3

Kimi K3 launched July 16, 2026 with 2.8 trillion total parameters (280B active) and a 1,048,576-token context window, the largest announced open-weight-class model to date. It targets teams doing agentic coding and long-document analysis who want an open alternative to GPT-5.6 Sol and Claude Opus 4.8.

Kimi K3 is Moonshot AI's 2.8-trillion-parameter mixture-of-experts model released July 16, 2026, scoring 93.5% on GPQA Diamond, the highest published open-weight result. It pairs a 1M-token context window with native vision input and Kimi Delta Attention for long-horizon coding and reasoning.

Provider: Moonshot AI · Family: Kimi

More about Moonshot AI on HokAI

Context window: 1,048,576 tokens · Max output: 128,000

Input modalities: text, image · Output: text, tool-calls, code

About Kimi K3

Kimi K3 is a 2.8-trillion-parameter mixture-of-experts language model released by Chinese AI lab Moonshot AI on July 16, 2026. It uses 896 experts with 16 active per token (roughly 280B active parameters) and Kimi Delta Attention, a hybrid linear attention mechanism paired with Attention Residuals designed to keep recall stable across very long context. K3 succeeds the Kimi K2.7 Code line and is positioned as Moonshot's frontier release for agentic coding, long-horizon knowledge work, and multimodal reasoning, competing directly with GPT-5.6 Sol, Claude Opus 4.8, and Claude Fable 5. On benchmarks, K3 posted 93.5% on GPQA Diamond, the highest published score among open-weight models and ahead of Opus 4.8's 91.0%. On agentic coding evals it scored 67.5% on DeepSWE using the KimiCode harness (67.3% under the mini-SWE-agent harness) and 88.3% on Terminal-Bench 2.1, half a point behind GPT-5.6 Sol and ahead of every other open or closed model tested. It also leads Program Bench at 77.8% and SWE Marathon at 42.0%. Independent evaluator Artificial Analysis put its blended Intelligence Index at 57, with a measured cost of $0.94 per task, cheaper than GPT-5.6 Sol ($1.04) and roughly half of Opus 4.8 ($1.80), though well above open peers like GLM-5.2 ($0.32) and DeepSeek V4 Pro ($0.04). The context window is 1,048,576 tokens (1M) with a 128,000-token max output. Kimi Delta Attention is built to hold recall across the full window rather than degrading past 100K the way sliding-window architectures do, matching GPT-5.5's context length and exceeding the 200K windows on GPT-5.6 Sol, Terra, and Luna. K3 accepts text and image input and produces text, tool-calls, and code output. It supports native vision (MMMU-Pro 81.6%, MathVision 97.8%), function calling, structured output, and code execution, but has no native audio or video input/output, so voice and video pipelines need a separate model. Reasoning modes span none, low, medium, high, xhigh, and max, though Moonshot has not published exact token budgets for each tier. Hosted API pricing on platform.moonshot.ai is $3.00 per million input tokens, $0.30 per million cached-hit input tokens (a 90% discount), and $15.00 per million output tokens, with no long-context surcharge across the full 1M window. The model is also available identically priced on OpenRouter as moonshotai/kimi-k3. A full 1M-token context analysis run locally on 8x H100 80GB costs roughly $32 in compute time; a 500K-token agentic coding session against the hosted API runs about $1.25. K3 is live today through kimi.com, the Kimi iOS app, Kimi Code, and the API at https://api.moonshot.ai/v1 (model ID kimi-k3). Moonshot has committed to releasing the weights under a Modified MIT-style license, the same license used for K2.5/K2.6/K2.7 Code, by July 27, 2026, but as of this model's publication the weights are not yet live on Moonshot's Hugging Face org, which still tops out at K2.7 Code. Until the weights land, self-hosting is not possible and K3 is open-weight in commitment only, not in practice. Once released, full FP16 inference needs roughly 5.6TB of VRAM (H100 80GB x8), INT8 needs about 2.8TB (H100 80GB x2), and INT4 quantization (AWQ/GPTQ/GGUF) brings it down to around 1.4TB, usable on 4x RTX 4090 24GB. Moonshot uses RLHF alignment with an internal safety team and a documented content-filter and moderation API. Refusals cover standard categories: CSAM, weapons, self-harm, PII exposure, and election interference. The independently measured HarmBench refusal rate is 88%. The more notable safety signal is a hallucination rate of 51% on Artificial Analysis's AA-Omniscience benchmark, up from 39% on K2.6, meaning K3 answers more confidently while fabricating more often, so outputs on factual tasks should be verified or grounded with RAG. K3 is the strongest choice for teams that want an open-weight frontier model for long-context agentic coding, GPQA-class reasoning, or fine-tuning research, and who can tolerate Chinese-first documentation and a pending weights release. Teams needing verified factual reliability, native audio/video, or Western enterprise compliance certifications (no SOC 2, HIPAA, or ISO 27001 disclosed) are better served by Opus 4.8 or GPT-5.6 Sol. Training data includes public web text, licensed datasets, synthetic reasoning traces, and Moonshot's internal Chinese/English corpora; Moonshot has not published an exact training cutoff date. Hosted API inputs are retained 30 days for abuse monitoring under Moonshot's standard policy; local inference after the weight release keeps all data on the user's own infrastructure.

Pricing

Hosted API (platform.moonshot.ai / api.moonshot.ai): $3.00 per 1M input tokens, $0.30 per 1M cached-hit input tokens (90% discount), $15.00 per 1M output tokens, flat across the full 1M context with no long-context surcharge. Identical pricing on OpenRouter as moonshotai/kimi-k3. Once open weights ship (expected by 2026-07-27), local inference is hardware cost only under the Modified MIT license.

Key Features

• World's Largest Announced Open-Weight-Class MoE: 2.8 trillion total parameters across 896 experts, 16 active per token, the largest open-weight-class model announced to date, releasing under a Modified MIT license by July 27, 2026.
• 1M Context Window with Delta Attention: Kimi Delta Attention maintains high recall across the full 1,048,576-token window, matching GPT-5.5 and exceeding the 200K windows on the GPT-5.6 family.
• Native Vision on an Open Model: Multimodal text and image input with strong vision benchmarks (MMMU-Pro 81.6%, MathVision 97.8%), a rare combination among open-weight-class models.
• Open-Weight-Class GPQA Diamond Leader: 93.5% GPQA Diamond is the highest published open-weight score, ahead of Opus 4.8's 91.0% and roughly matching GPT-5.5's 93.5%.
• Cost-Efficient Hosted Inference: Artificial Analysis measured a $0.94 blended cost per task on its Intelligence Index, undercutting GPT-5.6 Sol ($1.04) and about half of Opus 4.8 ($1.80).

Pros

• Highest published open-weight GPQA Diamond score at 93.5%, beating Opus 4.8's 91.0%.
• Terminal-Bench 2.1 score of 88.3%, half a point behind GPT-5.6 Sol and ahead of every other open or closed model tested.
• Program Bench 77.8% and SWE Marathon 42.0%, leading the field including proprietary flagships.
• 1M-token context window with high recall, matching GPT-5.5 and exceeding GPT-5.6 Sol's 200K.
• $0.94 blended Intelligence Index cost per task, cheaper than GPT-5.6 Sol and about half of Opus 4.8.

Cons

• Hallucination rate rose to 51% on AA-Omniscience versus 39% for predecessor K2.6.
• Open weights committed under Modified MIT but not yet published as of 2026-07-19; expected by 2026-07-27.
• 2.8T parameters need roughly 5.6TB VRAM for full FP16 local inference once weights ship.
• No native audio or video input/output, vision plus text only.
• No SOC 2, ISO 27001, or HIPAA disclosed, documentation and community are Chinese-first.

Benchmarks

• math: 94
• mmlu: 95
• mmlu pro: 87
• aime 2025: 89
• arc agi 2: 32
• humaneval: 92
• live bench: 68
• lmarena elo: 1380
• gpqa diamond: 93.5
• lmarena rank: 8
• aider polyglot: 77.8
• swe bench verified: 67.5
• humanitys last exam: 43.5
• artificial analysis intelligence index: 57
• artificial analysis price blended per m: 0.94

Visit Kimi K3 Official Page