Per-tensor, imatrix-weighted mixed-bit quantization of deepseek-ai/DeepSeek-V4.1-Flash-1017 (revision 1bae0e1088993b9b82de08f59de5be5a34e42170) sized to run on a 256 GB Apple Silicon Mac (such as an M5 Ultra 256 GB) with a full 1,048,576-token context window, the DSpark MTP drafter, and the vision tower loaded.
216.14 GiB, 551.88B weights; 3.36 bpw / 215.56 GiB across the 46,429 re-quantizable matrices, down from 4.33 bpw in the release checkpoint)209.60 GiB, 543.58B weights) — 3.77 bpw active when weighted by calibration routing frequency, because 56.8% of routed token traffic hits experts kept in lossless release MXFP4.safetensors shards (224.44 GiB of tensors, 240.98 GB on disk with 256-byte tensor alignment, including the unwired 7.39 GiB DSpark drafter and 0.90 GiB vision tower)unsloth/DeepSeek-V4.1-Flash-1017-GGUF (imatrix_unsloth.gguf_file).bitmap.json): Multiple-choice knapsack optimization over exact GPU-measured imatrix-weighted relative error across all 46,429 re-quantizable matrices under a 215.555 GiB wired-weight budget.weight = scale * q + bias with BF16 scales and biases at group size 64 or 32), initialized with min-max and refined with two imatrix-weighted least-squares refits. All keep tensors are byte-identical copies from deepseek-ai/DeepSeek-V4.1-Flash-1017.tensorfold-dsv41-mixedbit v1): Documented in tensorfold_mixedbit.json and config.json (quantization_config.quant_method = "tensorfold_mixedbit"). Every tensor starts at a 256-byte-aligned offset for zero-copy Metal buffer mapping, and shards break cleanly before mtp.* and vision.* so drafter and vision weights are not wired into the backbone residency set.| Component | Weights | Release Format | Mixed-Bit Assignment | Avg bpw | Size (GiB) |
|---|---|---|---|---|---|
Routed experts (w1, w2, w3) |
543.58B | MXFP4 (4.25 bpw) | 12,795 keep MXFP4 · 15,030 a3g64 · 18,246 a2g64 · 9 a2g32 |
3.312 (3.775 active) | 209.60 |
| Attention | 5.13B | FP8 / BF16 | 100 keep · 120 a6g64 · 4 a8g64 · 3 a5g64 |
6.620 | 3.95 |
| Shared experts | 1.42B | FP8 | 116 keep FP8 · 4 a6g64 |
7.958 | 1.31 |
embed.weight + head.weight |
1.32B | BF16 | 2 a4g64 |
4.500 | 0.69 |
Kept backbone (Engram dense, routers, norms, hc_*) |
0.43B | FP8 / BF16 / F32 | keep |
11.648 | 0.59 |
DSpark drafter (mtp.*, unwired) |
14.23B | MXFP4 / FP8 / BF16 / F32 | keep |
4.461 | 7.39 |
| Vision tower (unwired) | 0.49B | BF16 | keep |
16.000 | 0.90 |
.engram.embed.*) are unchanged from deepseek-ai/DeepSeek-V4.1-Flash-1017 and are not duplicated here ("engram_tables": "release" in tensorfold_mixedbit.json). Set TF_DSV41_ENGRAM_RELEASE to the local deepseek-ai/DeepSeek-V4.1-Flash-1017 directory and TF_DSV41_ENGRAM_DIR to the packed Engram directory.sudo sysctl iogpu.wired_limit_mb=249036) and pass DSV41_LOAD_LIMIT_GB=235 when serving so the 216.5 GiB backbone residency set stays wired above the default 70%-of-RAM threshold.173/200) vs. 89.0% (178/200) release; single-tool function calling 20/20; long-context needle recall 9/9 across 32K, 131K, and 500K, plus verified at 1,011,610 tokens on an M5 Ultra 256 GB.2.3 s warm load).Inherits the MIT license of deepseek-ai/DeepSeek-V4.1-Flash-1017. Importance matrix calibration statistics derived from unsloth/DeepSeek-V4.1-Flash-1017-GGUF.
model-00032-of-00047.safetensors 5.4 GBmodel-00042-of-00047.safetensors 5.4 GBmodel-00023-of-00047.safetensors 5.4 GBmodel-00029-of-00047.safetensors 5.4 GBmodel-00022-of-00047.safetensors 5.4 GBmodel-00038-of-00047.safetensors 5.4 GBmodel-00033-of-00047.safetensors 5.4 GBmodel-00034-of-00047.safetensors 5.4 GBmodel-00008-of-00047.safetensors 5.4 GBmodel-00007-of-00047.safetensors 5.4 GBmodel-00006-of-00047.safetensors 5.4 GBmodel-00012-of-00047.safetensors 5.4 GBmodel-00043-of-00047.safetensors 5.4 GBmodel-00040-of-00047.safetensors 5.4 GBmodel-00041-of-00047.safetensors 5.4 GBmodel-00037-of-00047.safetensors 5.4 GBmodel-00004-of-00047.safetensors 5.4 GBmodel-00025-of-00047.safetensors 5.4 GBmodel-00013-of-00047.safetensors 5.4 GBmodel-00039-of-00047.safetensors 5.4 GBmodel-00018-of-00047.safetensors 5.4 GBmodel-00002-of-00047.safetensors 5.4 GBmodel-00019-of-00047.safetensors 5.4 GBmodel-00031-of-00047.safetensors 5.4 GBmodel-00028-of-00047.safetensors 5.4 GBmodel-00011-of-00047.safetensors 5.4 GBmodel-00030-of-00047.safetensors 5.4 GBmodel-00009-of-00047.safetensors 5.4 GBmodel-00003-of-00047.safetensors 5.4 GBmodel-00027-of-00047.safetensors 5.4 GBmodel-00035-of-00047.safetensors 5.4 GBmodel-00005-of-00047.safetensors 5.4 GBmodel-00014-of-00047.safetensors 5.4 GBmodel-00036-of-00047.safetensors 5.4 GBmodel-00045-of-00047.safetensors 5.4 GBmodel-00015-of-00047.safetensors 5.4 GBmodel-00026-of-00047.safetensors 5.4 GBmodel-00017-of-00047.safetensors 5.4 GBmodel-00016-of-00047.safetensors 5.4 GBmodel-00001-of-00047.safetensors 5.4 GBmodel-00020-of-00047.safetensors 5.4 GBmodel-00021-of-00047.safetensors 5.4 GBmodel-00010-of-00047.safetensors 5.4 GBmodel-00024-of-00047.safetensors 5.4 GBmodel-00046-of-00047.safetensors 2.6 GBmodel-00044-of-00047.safetensors 1.3 GBmodel-00047-of-00047.safetensors 971 MB