Up to 9.5 tokens/sec so far on a 256 GiB M3 Ultra Mac Studio in short, text-only custom-runtime tests.
Original model · DeepSeek · Apple MLX · MIT licence
Experimental weights and standalone runtime, not a drop-in oMLX release. The custom text adapter is included in runtime/. Stock oMLX and standard MLX loaders have not been validated for this source-layout checkpoint. Read the memory requirements and limitations before downloading.
| Item | Value |
|---|---|
| Base model | deepseek-ai/DeepSeek-V4.1-Flash |
| Format | Source-layout MLX safetensors |
| Quantisation | Affine 2-bit, group size 64; selected non-matrix parameters remain higher precision |
| Source precision | Official mixed FP4/FP8 checkpoint, with BF16/F32 components; not a BF16-source conversion |
| Conversion type | Standard requantisation, not calibrated oQ |
| Weight payload | 238.80 GB / 222.40 GiB, plus headers and tokenizer/config files |
| Tensors / shards | 143,982 tensors / 48 shards |
| Native MTP | All three DSpark stages retained; 3,584 converted MTP tensors |
| Architecture | deepseek_v41, with Engram conditional memory |
| Context | Upstream config specifies 1,048,576 tokens; long context is not validated here |
Measured on 2026-09-10 with MLX 0.32.0, greedy decoding and the upstream chat encoder in chat mode. The custom adapter keeps the text backbone in RAM, leaves Engram tables on SSD, and can cache frequently used Engram rows in process memory.
| Short diagnostic | Output tokens | Timed decode steps | Non-MTP, warm Engram cache | Native MTP, serial verification |
|---|---|---|---|---|
| Arithmetic, 14-token prompt | 9 including EOS | 8 | 9.46 tokens/sec | 8.54 tokens/sec |
| Python function, 16-token prompt | 24, stopped at output cap | 23 | 8.80 tokens/sec | 8.17 tokens/sec |
The headline 9.5 tokens/sec rounds the 9.46 result above. It is a short observed result, not a sustained-speed guarantee, a broad benchmark, or an oMLX result. Warm cache means the same prompt was run previously; real conversations may have different cache hit rates. Initial model loading and prompt processing are excluded from decode speed.
The optimized non-MTP adapter also produced 8.62–8.97 tokens/sec in a separate repeated two-prompt comparison without the explicit Engram row cache. Native MTP produced identical greedy outputs on both diagnostics but was slower with serial verification; accelerated block verification is not validated. Observed accepted draft tokens averaged 1.33 and 1.00 per block, respectively, with output limits truncating some proposals.
The measured text backbone occupied 160.88 GiB; 57.22 GiB of Engram tables stayed on SSD. The native MTP weights add 4.15 GiB when loaded. A combined benchmark session reached 166.69 GiB peak process RSS, excluding filesystem-cache memory and other system use.
These observations apply to the custom offload strategy on a 256 GiB Mac, not to loading every shard into GPU memory at once. Keep substantial room for macOS, temporary allocations and context growth. Cold setup took approximately three minutes in the development environment; storage performance will change that figure.
Prompt, using the upstream chat encoder and greedy decoding:
What is 2+2? Answer briefly.
Actual output, excluding the EOS marker:
2 plus 2 equals 4.
The coding diagnostic checked a 24-token prefix only, not a completed or executed Python function. Neither test establishes broad reasoning, coding or chat quality.
hf download Vontra/DeepSeek-V4.1-Flash-MLX-2bit-MTP --local-dir DeepSeek-V4.1-Flash-MLX-2bit-MTP
The download includes the standalone runtime, pinned dependencies and tests. On a 256 GiB Apple silicon Mac, from the downloaded directory:
python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install -r runtime/requirements.txt
python -m unittest discover -s runtime -p 'test_*.py'
python runtime/generate.py --model . --resident-backbone --prompt "What is 2+2? Answer briefly." --max-tokens 16 --repeat 2
The second run reuses the model and Engram row cache. Add --mtp to test native DSpark with serial verification, which is currently slower. The command releases memory when it exits; no server is started. See the runtime guide for the 128-token diagnostic context cap and other restrictions. This does not install support into oMLX.
For the upstream architecture, evaluations, intended use and limitations, see the original model card.
The original model is released under the MIT licence, included as LICENSE. Model design, training and upstream documentation belong to DeepSeek and its contributors; this is an independent community conversion by Vontra, not an official DeepSeek release.
model-00048-of-00048.safetensors 31 GBmodel-00047-of-00048.safetensors 31 GBmodel-00017-of-00048.safetensors 4.3 GBmodel-00005-of-00048.safetensors 4.3 GBmodel-00011-of-00048.safetensors 4.3 GBmodel-00023-of-00048.safetensors 4.3 GBmodel-00027-of-00048.safetensors 4.3 GBmodel-00031-of-00048.safetensors 4.3 GBmodel-00035-of-00048.safetensors 4.3 GBmodel-00039-of-00048.safetensors 4.3 GBmodel-00013-of-00048.safetensors 4.3 GBmodel-00014-of-00048.safetensors 4.3 GBmodel-00015-of-00048.safetensors 4.3 GBmodel-00016-of-00048.safetensors 4.3 GBmodel-00018-of-00048.safetensors 4.3 GBmodel-00019-of-00048.safetensors 4.3 GBmodel-00020-of-00048.safetensors 4.3 GBmodel-00021-of-00048.safetensors 4.3 GBmodel-00022-of-00048.safetensors 4.3 GBmodel-00024-of-00048.safetensors 4.3 GBmodel-00025-of-00048.safetensors 4.3 GBmodel-00026-of-00048.safetensors 4.3 GBmodel-00028-of-00048.safetensors 4.3 GBmodel-00029-of-00048.safetensors 4.3 GBmodel-00030-of-00048.safetensors 4.3 GBmodel-00032-of-00048.safetensors 4.3 GBmodel-00033-of-00048.safetensors 4.3 GBmodel-00034-of-00048.safetensors 4.3 GBmodel-00036-of-00048.safetensors 4.3 GBmodel-00037-of-00048.safetensors 4.3 GBmodel-00038-of-00048.safetensors 4.3 GBmodel-00040-of-00048.safetensors 4.3 GBmodel-00041-of-00048.safetensors 4.3 GBmodel-00042-of-00048.safetensors 4.3 GBmodel-00003-of-00048.safetensors 4.3 GBmodel-00004-of-00048.safetensors 4.3 GBmodel-00006-of-00048.safetensors 4.3 GBmodel-00007-of-00048.safetensors 4.3 GBmodel-00008-of-00048.safetensors 4.3 GBmodel-00009-of-00048.safetensors 4.3 GBmodel-00010-of-00048.safetensors 4.3 GBmodel-00012-of-00048.safetensors 4.3 GBmodel-00044-of-00048.safetensors 1.5 GBmodel-00046-of-00048.safetensors 1.5 GBmodel-00045-of-00048.safetensors 1.5 GBmodel-00002-of-00048.safetensors 207 MBmodel-00043-of-00048.safetensors 207 MBmodel-00001-of-00048.safetensors 153 MBdeepseek-ai/DeepSeek-V4.1-FlashHave 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 0xSojalSec/DeepSeek-V4.1-Flash-MLX-MAC ./model-folder