Mixed-precision (UVMAX) quantization of Qwen/Qwen3.8-2.4T-A95B, converted with mlx-lm from the bf16 release.
UVMAX assigns bit widths per tensor class instead of quantizing uniformly.
| tensor class | precision | parameters | size | share |
|---|---|---|---|---|
| Expert FFN gate/up | 2-bit, group 128 | 1,580B | 414 GiB | 55.2% |
| Expert FFN down | 3-bit, group 128 | 790B | 299 GiB | 39.9% |
| Attention + gated delta net | 6-bit, group 64 | 40B | 30.2 GiB | 4.0% |
| Shared experts | 8-bit, group 64 | 4.6B | 4.6 GiB | 0.6% |
| Embeddings, LM head | 4-bit, group 64 | 4.1B | 2.1 GiB | 0.3% |
| Routers | 8-bit, group 64 | 0.4B | 0.4 GiB | <0.1% |
| Norms, convolutions, gates | bf16 / fp32 | 0.01B | — | <0.1% |
| total | 2.66 bits/weight | 2.42T | 750 GiB |
Teacher-forced against the bf16 release on identical tokens, 48 windows of
1025 tokens. KLD is KL(bf16 ‖ UVMAX) over the full output distribution and
is corpus-specific.
| corpus | ppl bf16 | ppl UVMAX | ratio | mean KLD | median KLD | top-1 agreement |
|---|---|---|---|---|---|---|
| Linux kernel C | 1.300 | 1.406 | 1.08× | 0.099 | 0.0008 | 94.9% |
| XNU kernel C | 2.122 | 2.272 | 1.07× | 0.106 | 0.0061 | 91.7% |
| JavaScriptCore C++ | 2.059 | 2.184 | 1.06× | 0.102 | 0.0031 | 92.1% |
| English prose | 2.684 | 3.188 | 1.19× | 0.225 | 0.0528 | 84.8% |
| all | 1.976 | 2.172 | 1.10× | 0.133 | 0.0062 | 90.9% |
2× M3 Ultra (512 GB), tensor parallelism over Thunderbolt RDMA, batch size 1, peak memory per node.
| prompt tokens | prefill tok/s | generation tok/s | peak memory |
|---|---|---|---|
| 4k | 176 | 14.2 | 385 GiB |
| 16k | 171 | 13.7 | 385 GiB |
| 32k | 162 | 13.1 | 386 GiB |
| 64k | 142 | 12.3 | 391 GiB |
| 128k | 113 | 10.7 | 402 GiB |
Requires mlx-lm 0.31.3 or later.
pip install mlx-lm
The model does not fit on one 512 GB machine. Serve it across two with an MLX distributed hostfile:
mlx.launch --backend jaccl --hostfile hosts.json --env MLX_METAL_FAST_SYNCH=1 -- \
mlx_lm.server --model kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX
Sampling follows the base model: temperature 1.0, top-p 0.95, top-k 20.
model-00028-of-00185.safetensors 4.8 GBmodel-00036-of-00185.safetensors 4.8 GBmodel-00078-of-00185.safetensors 4.8 GBmodel-00100-of-00185.safetensors 4.8 GBmodel-00160-of-00185.safetensors 4.8 GBmodel-00022-of-00185.safetensors 4.8 GBmodel-00034-of-00185.safetensors 4.8 GBmodel-00042-of-00185.safetensors 4.8 GBmodel-00044-of-00185.safetensors 4.8 GBmodel-00046-of-00185.safetensors 4.8 GBmodel-00054-of-00185.safetensors 4.8 GBmodel-00056-of-00185.safetensors 4.8 GBmodel-00064-of-00185.safetensors 4.8 GBmodel-00066-of-00185.safetensors 4.8 GBmodel-00068-of-00185.safetensors 4.8 GBmodel-00070-of-00185.safetensors 4.8 GBmodel-00072-of-00185.safetensors 4.8 GBmodel-00074-of-00185.safetensors 4.8 GBmodel-00084-of-00185.safetensors 4.8 GBmodel-00088-of-00185.safetensors 4.8 GBmodel-00096-of-00185.safetensors 4.8 GBmodel-00104-of-00185.safetensors 4.8 GBmodel-00108-of-00185.safetensors 4.8 GBmodel-00118-of-00185.safetensors 4.8 GBmodel-00126-of-00185.safetensors 4.8 GBmodel-00130-of-00185.safetensors 4.8 GBmodel-00134-of-00185.safetensors 4.8 GBmodel-00138-of-00185.safetensors 4.8 GBmodel-00140-of-00185.safetensors 4.8 GBmodel-00152-of-00185.safetensors 4.8 GBmodel-00164-of-00185.safetensors 4.8 GBmodel-00168-of-00185.safetensors 4.8 GBmodel-00182-of-00185.safetensors 4.8 GBmodel-00040-of-00185.safetensors 4.8 GBmodel-00058-of-00185.safetensors 4.8 GBmodel-00062-of-00185.safetensors 4.8 GBmodel-00076-of-00185.safetensors 4.8 GBmodel-00086-of-00185.safetensors 4.8 GBmodel-00090-of-00185.safetensors 4.8 GBmodel-00092-of-00185.safetensors 4.8 GBmodel-00106-of-00185.safetensors 4.8 GBmodel-00110-of-00185.safetensors 4.8 GBmodel-00112-of-00185.safetensors 4.8 GBmodel-00116-of-00185.safetensors 4.8 GBmodel-00120-of-00185.safetensors 4.8 GBmodel-00128-of-00185.safetensors 4.8 GBmodel-00136-of-00185.safetensors 4.8 GBmodel-00142-of-00185.safetensors 4.8 GBmodel-00154-of-00185.safetensors 4.8 GBmodel-00162-of-00185.safetensors 4.8 GBmodel-00166-of-00185.safetensors 4.8 GBmodel-00174-of-00185.safetensors 4.8 GBmodel-00176-of-00185.safetensors 4.8 GBmodel-00178-of-00185.safetensors 4.8 GBmodel-00180-of-00185.safetensors 4.8 GBmodel-00016-of-00185.safetensors 4.8 GBmodel-00024-of-00185.safetensors 4.8 GBmodel-00026-of-00185.safetensors 4.8 GBmodel-00032-of-00185.safetensors 4.8 GBmodel-00060-of-00185.safetensors 4.8 GBmodel-00114-of-00185.safetensors 4.8 GBmodel-00122-of-00185.safetensors 4.8 GBmodel-00144-of-00185.safetensors 4.8 GBmodel-00146-of-00185.safetensors 4.8 GBmodel-00148-of-00185.safetensors 4.8 GBmodel-00158-of-00185.safetensors 4.8 GBmodel-00172-of-00185.safetensors 4.8 GBmodel-00004-of-00185.safetensors 4.8 GBmodel-00010-of-00185.safetensors 4.8 GBmodel-00030-of-00185.safetensors 4.8 GBmodel-00038-of-00185.safetensors 4.8 GBmodel-00048-of-00185.safetensors 4.8 GBmodel-00050-of-00185.safetensors 4.8 GBmodel-00052-of-00185.safetensors 4.8 GBmodel-00080-of-00185.safetensors 4.8 GBmodel-00082-of-00185.safetensors 4.8 GBmodel-00094-of-00185.safetensors 4.8 GBmodel-00098-of-00185.safetensors 4.8 GBmodel-00102-of-00185.safetensors 4.8 GBmodel-00124-of-00185.safetensors 4.8 GBmodel-00132-of-00185.safetensors 4.8 GBmodel-00150-of-00185.safetensors 4.8 GBmodel-00156-of-00185.safetensors 4.8 GBmodel-00170-of-00185.safetensors 4.8 GBmodel-00184-of-00185.safetensors 4.8 GBmodel-00006-of-00185.safetensors 4.8 GBmodel-00008-of-00185.safetensors 4.8 GBmodel-00014-of-00185.safetensors 4.8 GBmodel-00012-of-00185.safetensors 4.8 GBmodel-00018-of-00185.safetensors 4.8 GBmodel-00020-of-00185.safetensors 4.8 GBmodel-00185-of-00185.safetensors 4.7 GBmodel-00001-of-00185.safetensors 3.9 GBmodel-00061-of-00185.safetensors 3.9 GBmodel-00181-of-00185.safetensors 3.9 GBmodel-00169-of-00185.safetensors 3.9 GBmodel-00051-of-00185.safetensors 3.9 GBmodel-00099-of-00185.safetensors 3.9 GBmodel-00155-of-00185.safetensors 3.9 GBmodel-00123-of-00185.safetensors 3.9 GBmodel-00117-of-00185.safetensors 3.9 GBmodel-00139-of-00185.safetensors 3.9 GBmodel-00153-of-00185.safetensors 3.9 GBmodel-00161-of-00185.safetensors 3.9 GBmodel-00037-of-00185.safetensors 3.9 GBmodel-00075-of-00185.safetensors 3.9 GBmodel-00129-of-00185.safetensors 3.9 GBmodel-00125-of-00185.safetensors 3.9 GBmodel-00147-of-00185.safetensors 3.9 GBmodel-00177-of-00185.safetensors 3.9 GBmodel-00163-of-00185.safetensors 3.9 GBmodel-00137-of-00185.safetensors 3.9 GBmodel-00043-of-00185.safetensors 3.9 GBmodel-00105-of-00185.safetensors 3.9 GBmodel-00053-of-00185.safetensors 3.9 GBmodel-00073-of-00185.safetensors 3.9 GBmodel-00077-of-00185.safetensors 3.9 GBmodel-00089-of-00185.safetensors 3.9 GBmodel-00093-of-00185.safetensors 3.9 GBmodel-00133-of-00185.safetensors 3.9 GBmodel-00141-of-00185.safetensors 3.9 GBmodel-00149-of-00185.safetensors 3.9 GBmodel-00009-of-00185.safetensors 3.9 GBmodel-00107-of-00185.safetensors 3.9 GBmodel-00121-of-00185.safetensors 3.9 GBmodel-00045-of-00185.safetensors 3.9 GBmodel-00083-of-00185.safetensors 3.9 GBmodel-00173-of-00185.safetensors 3.9 GBmodel-00025-of-00185.safetensors 3.9 GBmodel-00101-of-00185.safetensors 3.9 GBmodel-00033-of-00185.safetensors 3.9 GBmodel-00097-of-00185.safetensors 3.9 GBmodel-00171-of-00185.safetensors 3.9 GBmodel-00059-of-00185.safetensors 3.9 GBmodel-00091-of-00185.safetensors 3.9 GBmodel-00165-of-00185.safetensors 3.9 GBmodel-00029-of-00185.safetensors 3.9 GBmodel-00065-of-00185.safetensors 3.9 GBmodel-00085-of-00185.safetensors 3.9 GBmodel-00131-of-00185.safetensors 3.9 GBmodel-00049-of-00185.safetensors 3.9 GBmodel-00027-of-00185.safetensors 3.9 GBmodel-00067-of-00185.safetensors 3.9 GBmodel-00081-of-00185.safetensors 3.9 GBmodel-00113-of-00185.safetensors 3.9 GBmodel-00157-of-00185.safetensors 3.9 GBmodel-00179-of-00185.safetensors 3.9 GBmodel-00057-of-00185.safetensors 3.9 GBmodel-00069-of-00185.safetensors 3.9 GBmodel-00115-of-00185.safetensors 3.9 GBmodel-00021-of-00185.safetensors 3.9 GBmodel-00003-of-00185.safetensors 3.9 GBmodel-00005-of-00185.safetensors 3.9 GBmodel-00013-of-00185.safetensors 3.9 GBmodel-00017-of-00185.safetensors 3.9 GBmodel-00145-of-00185.safetensors 3.9 GBmodel-00109-of-00185.safetensors 3.9 GBmodel-00041-of-00185.safetensors 3.9 GBmodel-00011-of-00185.safetensors 3.9 GBmodel-00019-of-00185.safetensors 3.9 GBmodel-00035-of-00185.safetensors 3.9 GBmodel-00055-of-00185.safetensors 3.9 GBmodel-00183-of-00185.safetensors 3.9 GBmodel-00023-of-00185.safetensors 3.9 GBmodel-00095-of-00185.safetensors 3.9 GBmodel-00135-of-00185.safetensors 3.9 GBmodel-00031-of-00185.safetensors 3.9 GBmodel-00047-of-00185.safetensors 3.9 GBmodel-00063-of-00185.safetensors 3.9 GBmodel-00119-of-00185.safetensors 3.9 GBmodel-00159-of-00185.safetensors 3.9 GBmodel-00111-of-00185.safetensors 3.9 GBmodel-00143-of-00185.safetensors 3.9 GBmodel-00039-of-00185.safetensors 3.9 GBmodel-00103-of-00185.safetensors 3.9 GBmodel-00167-of-00185.safetensors 3.9 GBmodel-00151-of-00185.safetensors 3.9 GBmodel-00071-of-00185.safetensors 3.9 GBmodel-00175-of-00185.safetensors 3.9 GBmodel-00079-of-00185.safetensors 3.9 GBmodel-00015-of-00185.safetensors 3.9 GBmodel-00087-of-00185.safetensors 3.9 GBmodel-00127-of-00185.safetensors 3.9 GBmodel-00007-of-00185.safetensors 3.9 GBmodel-00002-of-00185.safetensors 2.4 GBQwen/Qwen3.8-2.4T-A95BHave the files? Submit a torrent.
Submit a torrentHave the complete files and permission to share them? Replace the example folder path. The script checks the files, makes a torrent and submits its magnet and checksums. Keep your torrent client seeding. Get your community key from your settings, and read the script before running it.
curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX ./model-folder