Kimi K3 Weights Land on HuggingFace: 2.8T Open Frontier Model You Can Actually Download

The full Kimi K3 model weights went live on HuggingFace today, July 27, 2026. The release marks the first time a truly frontier-class open-weight model sits at the top of the leaderboard alongside GPT-5.6 Sol and Claude Fable 5.

The download is not small. At roughly 1.63TB spread across 96 safetensors files (~17GB each), this is a datacenter-scale model. The HN thread hit 1,037 points and 423 comments in hours.

What Moonshot Actually Released#

Kimi K3 is Moonshot AI's flagship model. The key specs from the HuggingFace model card and technical report:

Architecture. Mixture-of-Experts with 896 experts, 16 selected per token plus 2 shared. Hybrid attention: 69 Kimi Delta Attention layers + 24 Gated MLA layers. 93 layers total. Activation function: SiTU-GLU.

Scale. 2.8 trillion total parameters, 104 billion activated per token. Hidden dimension 7,168, MoE hidden dimension 3,072 per expert. Latent MoE dimension 3,584.

Context and vision. 1,048,576 token context window. Native multimodal support with MoonViT-V2 vision encoder (401M parameters). Supports text and image inputs.

Quantization. Native MXFP4 weights with MXFP8 activations, trained from the SFT stage onward. This is the first major open model to ship native 4-bit weights at this scale.

License. Custom Kimi K3 License. Free for most use, but requires a separate commercial agreement if you run a model-as-a-service business with over $20M in aggregate revenue over any 12 consecutive months. Also requires prominent "Kimi K3" branding if your product exceeds 100M MAU or $20M MRR.

What else shipped. Moonshot also open-sourced MoonEP (expert parallelism framework) and AgentEnv (agent evaluation environment), making the full inference stack reproducible.

Benchmarks at a Glance#

K3 goes toe-to-toe with the best closed models:

• GPQA Diamond: 93.5 (vs Fable 5 at 92.6, GPT-5.6 Sol at 94.1)
• ProgramBench: 77.8 (vs Fable 5 at 76.8, GPT-5.6 Sol at 77.6)
• Terminal-Bench 2.1: 88.3 (tops among open models)
• BrowseComp: 91.2 (beats both Fable 5 at 88.0 and GPT-5.6 Sol at 90.4)
• MCPMark-Verified: 94.5 (highest among all models listed)
• SWE-Marathon: 42.0 (leads the table; Fable 5 at 35.0)
• DeepSearchQA F1: 95.0 (beats Fable 5 at 94.2)

Moonshot published the full technical report with methodology notes, including harness choices and evaluation conditions.

What HN Is Saying#

The HN discussion at 423 comments split into several threads worth synthesizing.

Historical significance. Multiple commenters called this a watershed moment. "For the first time, an open-weights LLM is right at the top," wrote davidkunz. Another commenter (padolsey) compared it to publishing RSA source code on a t-shirt - once it is out there, there is no putting it back.

VRAM reality check. The model requires ~1.5TB of VRAM at native MXFP4. NitpickLawyer broke down the hosting math: "just at the limit of 8xB200s, but realistically you will need 16x for context/throughput optimisation." A 2-bit quant is already on HuggingFace at ~1TB. Most commenters agreed that individual developers are priced out - this is a datacenter model.

The license debate. The revenue-based license terms drew comparison to Meta's Llama approach. Moonshot's thresholds ($20M revenue for MAAS businesses) were seen as higher and more targeted at token resellers than Meta's earlier restrictions. Commenters noted the branding clause ($20M MRR) is smart marketing.

Market pricing cascade. One commenter (gorgmah) observed that GLM 5.2 prices dropped roughly 45% since its June 16 release, and the downward slope is continuing. The consensus: K3 entering the market will accelerate this trend. Several providers already serve K3 on OpenRouter, and price competition is expected to intensify.

Hardware gap frustration. KronisLV articulated a common sentiment: "most hardware to run LLMs on is shaped wrong for individuals." Prosumer GPUs with 128-256GB VRAM at reasonable TDP do not exist, leaving the free weights usable only by cloud providers and well-funded labs.

Security and fine-tuning. AISI benchmarks reportedly place K3 above GLM 5.2 on cybersecurity but still behind closed models. Commenters discussed whether LoRA, DPO, or distillation could produce consumer-friendly derivatives, with Unsloth and 1-bit quantizations mentioned as plausible paths.

Why This Matters for Developers#

Three takeaways from this release.

The open frontier is real now. Previous open models (Llama 4, Qwen 3, DeepSeek V4, GLM 5.2) were strong but visibly a tier below Claude and GPT on agentic workloads. K3 closes that gap. The vLLM, SGLang, and TokenSpeed ecosystems already support it. For teams building on open weights, there is now a credible option at the very top of the benchmark table.

Inference pricing keeps dropping. The GLM 5.2 pattern - 45% price decline in six weeks - is now the normal cycle for competitive open-weight releases. K3 entering the market means more downward pressure on high-end inference pricing. The comment thread repeatedly notes that this benefits every team running models at scale, not just those using K3 directly.

The license model is evolving. Moonshot's tiered approach - free below $20M, commercial terms above - targets token resellers specifically while leaving startups and most enterprises unaffected. It is a different philosophy than Apache 2.0 or the Llama licenses. If this becomes a template for future Chinese open-weight releases, the licensing landscape for open models will look very different in 12 months.

The weights are live on HuggingFace now. Whether you download them or just read the report, this is the first day the open-weights frontier truly matches the closed one.

Sources#

• HuggingFace model card: https://huggingface.co/moonshotai/Kimi-K3
• Kimi K3 Technical Report: https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf
• HN Discussion: https://news.ycombinator.com/item?id=49065752
• MoonEP framework: https://github.com/MoonshotAI/MoonEP
• AgentEnv: https://github.com/kvcache-ai/AgentEnv
• Kimi Code CLI: https://www.kimi.com/code