DeepSeek-V4.1-Flash-1017 — IQR hybrid
Radiance .rad of DeepSeek-V4.1-Flash-1017.
Weights are not in this repo yet.
Container size: 262.75 GiB (282,125,652,296 bytes). Lane-order copy of the same weights: 262.87 GiB (282,254,628,612 bytes).
| Container | 262.75 GiB / 282,125,652,296 bytes |
| Lane-order copy | 262.87 GiB / 282,254,628,612 bytes |
| Routed experts | IQ3R on layers 6, 7, 8, 9, 10, 12, 14, 16, 25, 28, 29, 31 (u3, tcq 16, mad, fwht128). Every other routed expert IQ2R (u2, tcq 15) |
| Drafter | layers 40–42 IQ3R |
| Attention | u6 / u7 |
| Engram table | iq4_nl, 103 GiB. Rebuilt from the tokenizer; the release does not ship it |
| Embeddings / output | q6_K |
| HumanEval | 144/164 |
| GSM8K | 198/200 |
| Decode | 74.5 tok/s (256 tokens, 4×32 GB RDNA4, TP4) |
| Prefill | 1106 tok/s (11,789 tokens) |
Other containers: uncensored UD-IQ1_S, 251.13 GiB · IQR IQ3R, 307.60 GiB planned
Index: https://huggingface.co/collections/bakon3/deepseek-v41-flash-radiance-6af1ddf468e089a2acfccafd