Kimi K3 Review (2026): 2.8T Open MoE, 1M Context, GPQA 93.5%, Apache 2.0
Kimi K3: 2.8T MoE open weights, 1M context, native vision, GPQA 93.5%, Terminal-Bench 88.3%. Apache 2.0. Intelligence Index 57. Honest review vs GPT-5.6, DeepSeek, GLM.
Kimi K3 is the world's first open 3T-class MoE (2.8T, Apache 2.0) with 1M context, native vision, GPQA 93.5% open-weight SOTA. Leads Program Bench, SWE Marathon, Terminal-Bench among open weights. Intelligence Index 57 = Opus 4.8 / GPT-5.5 tier. Hallucination 51% (up from 39%) — verify outputs. Best for open research, agentic coding SOTA, long-context RAG. Avoid for unverified production, consumer hardware, audio/video needs.
Kimi K3 is Moonshot AI's 2.8T MoE open-weight model (July 2026) with 1M context, native vision, and Apache 2.0 license. GPQA Diamond 93.5% (open-weight SOTA), Terminal-Bench 2.1 88.3% (0.5 pts behind GPT-5.6 Sol), DeepSWE 67.5%, Program Bench 77.8% (field-leading), SWE Marathon 42.0% (field-leading). Artificial Analysis Intelligence Index 57 — par with Opus 4.8, GPT-5.5. Agentic Elo 1,668 (GDPval v2) beats GPT-5.5, Opus 4.8, GLM-5.2. Hallucination rate 51% (AA-Omniscience) — up from 39% on K2.6. Local FP16 requires ~5.6TB VRAM; INT4 on 4x RTX 4090. Hosted API ~$0.50/$2.00 per 1M est. True open weights unlock fine-tuning, distillation, commercial deployment.
Provider: Moonshot AI · Family: Kimi
Context window: 1,000,000 tokens · Max output: 128,000
Input modalities: text, image · Output: text, tool-calls, code
Pricing
Open weights = free local inference (hardware cost only). Hosted API estimated $0.50 input / $2.00 output per 1M tokens (Moonshot platform). Independent eval (Artificial Analysis): $0.94 per task on Intelligence Index — competitive with GPT-5.6 Sol ($1.04), half of Opus 4.8 ($1.80). Well above open-weight peers (GLM-5.2 $0.32, DeepSeek V4 Pro $0.04).
Key Features
- World's First Open 3T-Class MoE: 2.8 trillion total parameters, 896 experts, 16 active per token — largest open-weight model with Apache 2.0 license.
- 1M Context Window with Delta Attention: Kimi Delta Attention architecture maintains high recall across full 1M tokens — matches GPT-5.5, exceeds GPT-5.6 family (200K).
- Native Vision on Open Weights: Multimodal input (text + image) with vision benchmarks (MMMU-Pro 81.6%, MathVision 97.8%) on open weights — rare combination.
- Open-Weight GPQA Diamond SOTA: 93.5% GPQA Diamond — highest published open-weight score, beats most proprietary flagships (Opus 4.8: 91.0%, GPT-5.5: 93.5%).
- Agentic Elo Leadership Among Open Weights: GDPval v2 Elo 1,668 — beats GLM-5.2, GPT-5.5, Opus 4.8; only Fable 5 higher. AutomationBench-AA 53% top spot.
Pros
- Highest published open-weight GPQA Diamond (93.5%) — beats all open and most proprietary models.
- Terminal-Bench 2.1 88.3% — just 0.5 points behind GPT-5.6 Sol, leads all other open and closed models.
- Program Bench 77.8% and SWE Marathon 42.0% — leads the entire field including proprietary flagships.
- True open-weight Apache 2.0 — commercial use, fine-tuning, local deployment, no license restrictions.
- 1M context window with high recall — matches GPT-5.5, exceeds GPT-5.6 Sol/Terra/Luna (200K).
- Agentic Elo 1,668 (GDPval v2) — beats GPT-5.5, Opus 4.8, GLM-5.2; only Fable 5 higher.
Cons
- Hallucination rate increased to 51% (AA-Omniscience) vs 39% for K2.6 — more fabrications despite higher accuracy.
- 2.8T parameters requires massive GPU memory for local inference — H100 80GB x8 for FP16; quantized versions needed for consumer hardware.
- Independent benchmark conditions vary (KimiCode vs Codex vs Claude Code harnesses) — not all head-to-heads use identical scaffolding.
- No native audio/video I/O — vision + text only; lags GPT-5.5/Gemini 3.1 on multimodal breadth.
- Hosted API pricing and availability outside China less established vs OpenAI/Anthropic gateways.
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: 65
- humanitys last exam: 43.5
- artificial analysis intelligence index: 57
- artificial analysis price blended per m: 0.94
Frequently Asked Questions
What is Kimi K3 and who built it?
Kimi K3 is Moonshot AI's third-generation frontier model, released July 16, 2026 as an open-weight model under Apache 2.0 license. It is a Mixture-of-Experts architecture with 2.8 trillion total parameters and 896 experts, activating 16 experts per token (estimated ~280B active parameters). The model supports a 1 million token context window and includes native vision capabilities built on Moonshot's Kimi Delta Attention architecture. Moonshot AI is a Beijing-based AI lab founded in March 2023 by Yang Zhilin and colleagues from Tsinghua University.
How much does Kimi K3 cost in 2026?
Kimi K3 is free to use locally — weights and code are Apache 2.0, permitting commercial use, fine-tuning, and deployment without license fees. Hardware cost only: H100 80GB x8 for FP16 (~$32/hr), H100 80GB x2 for INT8, 4x RTX 4090 24GB for INT4. Hosted API via Moonshot platform (platform.moonshot.cn) estimated at $0.50 input / $2.00 output per 1M tokens; exact international pricing TBD. Artificial Analysis independent eval: $0.94 per task on Intelligence Index — competitive with GPT-5.6 Sol ($1.04), half of Opus 4.8 ($1.80). Open-weight peers: GLM-5.2 $0.32, DeepSeek V4 Pro $0.04.
What are the main benchmarks for Kimi K3?
Key benchmarks (July 2026): GPQA Diamond 93.5% (open-weight SOTA, beats most proprietary); Terminal-Bench 2.1 88.3% (0.5 pts behind GPT-5.6 Sol, leads all others); DeepSWE 1.0 67.5%; Program Bench 77.8% (field-leading); SWE Marathon 42.0% (field-leading); FrontierSWE 81.2%; MMLU-Pro 87.0%; AIME 2025 89.0%; MMLU 95.0%; Math 94.0%; HumanEval 92.0%; LMArena Elo 1380 (rank 8); Artificial Analysis Intelligence Index 57 (par with Opus 4.8, GPT-5.5); Agentic Elo GDPval v2 1,668 (beats GPT-5.5, Opus 4.8, GLM-5.2). Hallucination: AA-Omniscience 51% (up from 39% on K2.6).
Is Kimi K3 truly open source?
Yes — Kimi K3 is released under the Apache 2.0 license, which is an OSI-approved permissive open-source license. This means: (1) Free commercial use — no license fees, no restrictions on revenue. (2) Free modification and fine-tuning — create derivatives, LoRAs, full fine-tunes. (3) Free redistribution — share weights, quantized versions, merged models. (4) No copyleft — derivatives can be proprietary. Weights available on Hugging Face (moonshotai/Kimi-K3), ModelScope, and GitHub. Code and training recipes also Apache 2.0. This is genuinely open-weight, not 'open weight with restrictions' like Llama Community License.
What hardware do I need to run Kimi K3 locally?
Full FP16: ~5.6TB VRAM (2.8T params × 2 bytes) — impractical (70x H100 80GB). INT8 quantization: ~2.8TB VRAM — 35x H100 80GB or 2x H100 80GB with tensor parallelism + CPU offloading. INT4 (AWQ/GPTQ): ~1.4TB VRAM — practical on 4x RTX 4090 24GB (96GB total) with offloading, or 2x H100 80GB. Recommended: H100 80GB x2 for INT8 quality; 4x RTX 4090 24GB for INT4 cost-efficiency. vLLM and SGLang support MoE tensor parallelism. Moonshot hosted API avoids hardware entirely. Quantized INT4 models available on Hugging Face/ModelScope.
How does Kimi K3 compare to GPT-5.6 Sol and DeepSeek V4 Pro?
Kimi K3 vs GPT-5.6 Sol: K3 open weights (Apache 2.0), 1M context (vs 200K), GPQA 93.5% (vs 94.1%), Terminal-Bench 88.3% (vs 88.8%), Intelligence Index 57 (vs 55), $0 license vs $5/$30 per 1M. Sol wins: hallucination 8.2% (vs 51%), ultra multi-agent, PTC, pro mode, hosted reliability. K3 vs DeepSeek V4 Pro: K3 higher Intelligence Index (57 vs ~45), 1M context (vs 128K), native vision, GPQA 93.5% (vs ~75%), Apache 2.0. V4 Pro wins: $0.04/task (vs $0.94), established international API, English-first docs, lower hardware needs (can run on fewer GPUs).
What is Kimi K3's hallucination problem?
Artificial Analysis AA-Omniscience Index shows Kimi K3 hallucination rate at 51%, up from 39% on Kimi K2.6. This means K3 fabricates answers more frequently even as its accuracy on known facts improves. The model is more confident but less reliable on uncertain or out-of-distribution queries. Mitigation: add explicit verification prompts ('Cite sources for each claim. Mark uncertain statements.'), use RAG for factual grounding, and treat all outputs as requiring verification for high-stakes use cases. This is a known regression from K2.6 that Moonshot has acknowledged in their model card.
Who is Kimi K3 best for and who should avoid it?
Best for: (1) Open-source LLM researchers — largest open MoE, Apache 2.0, fine-tunable. (2) Agentic coding benchmark teams — leads Program Bench, SWE Marathon, Terminal-Bench among open weights. (3) Long-context RAG architects — 1M context with Delta Attention high recall, matches GPT-5.5. (4) Cost-optimization at scale — $0.94/task Intelligence Index beats proprietary. Avoid for: (1) Unverified production — 51% hallucination rate. (2) Consumer hardware — INT4 still needs 4x RTX 4090. (3) Audio/video multimodal — vision+text only. (4) Teams without Chinese capability — docs/community primarily Chinese.