Within 16 days this summer, two labs put models of more than two trillion parameters on Hugging Face as downloadable weights: Moonshot AI released Kimi K3 on July 27, and the Qwen team at Alibaba uploaded Qwen3.8-2.4T-A95B between August 9 and August 12. Both downloads are ungated. What each lab opened, and on what conditions, differs more than the parameter counts suggest.
The two models side by side
| Kimi K3 | Qwen3.8-2.4T-A95B | |
|---|---|---|
| Weights released | July 27 | August 9–12 |
| Total / active | 2.8T / 104B | 2.4T / about 95B |
| Modalities | Text and images | Text only |
| Context | 1,048,576 tokens | 262,144 native, extensible to 1,010,000 |
| Checkpoint | 4-bit, about 1.56 TB | BF16, about 4.89 TB |
| License | Kimi K3 License | Qwen3.8-Max License |
Kimi K3 ships natively quantized: Moonshot trained it with quantization in mind from the fine-tuning stage, so the published weights are already 4-bit. Qwen publishes the full-precision checkpoint, and the smaller 4-bit builds come from third parties, not from Qwen.
Open, with conditions
Neither license is a standard open-source license. Both follow an MIT-style grant, free for commercial use, with added clauses. The Kimi K3 License requires a separate agreement with Moonshot before commercial use by any model-as-a-service operator whose revenue exceeds $20 million over 12 months, and products with more than 100 million monthly active users or more than $20 million in monthly revenue must display “Kimi K3.” The Qwen3.8-Max License sets its threshold at $50 million a year and extends it to “AI Work Assistant” businesses as well as model-as-a-service. Inside the same Qwen family, the smaller Qwen3.8-27B uses the standard Apache 2.0 license.
What it takes to run them
Downloadable does not mean runnable on modest hardware. According to the vLLM project’s recipes, Kimi K3 needs at least eight NVIDIA GB300 or eight AMD MI355X accelerators, with several nodes recommended in production. The Qwen flagship in full precision needs 24 B300 or 48 H200 GPUs; a single node of eight B300 fits only a 4-bit version.
Weights versus the hosted service
Some capabilities belong to each lab’s hosted service, not to the weights. On Moonshot’s platform, K3 costs $3.00 per million input tokens without cache, $0.30 with cache and $15.00 for output, with tools such as web search that run on the service side. Qwen’s showcase demos, from multi-day autonomous coding to long-video analysis, run on its hosted Qwen3.8-Max, and features such as deep research belong to its chat platform.
Both labs publish their own benchmark results. Moonshot documents its settings and harnesses and lists its own limitations, including a stated experience gap against Claude Fable 5 and GPT-5.6 Sol; Qwen says all its results come from its own testing. None were reproduced independently in the sources reviewed.
Related reading
- Kimi K3 repository — Moonshot AI
- Qwen3.8 repository — Qwen
- GLM-5.2 and the Chinese open-source bet on agent models — IA al Día