Run it on your Mac
curl -fsSL https://rapidmlx.com/install.sh | bash
rapid-mlx serve deepseek-v41-flash-reap-2bit
curl -s http://localhost:8000/v1/chat/completions -H 'Content-Type: application/json' -d '{"model":"default","messages":[{"role":"user","content":"Say hello"}]}'
Use it from OpenCode / Claude Code / Codex: rapid-mlx agents <name> --setup · Agent setup
Download the Desktop app · 256 GB unified memory required for this experimental 212.9 GB target and its DSpark sidecar.
DeepSeek V4.1 Flash · REAP · MLX 2-bit
Experimental research artifact — not yet a product-qualified Rapid-MLX model. It requires a 256 GiB Apple silicon Mac and the experimental loader that ships in Rapid-MLX 0.14.1. Standard model loaders are not supported. Vision and MTP are omitted.
This is a text-only, native packed MLX artifact produced from
Vontra/DeepSeek-V4.1-Flash-MLX-2bit-MTP,
itself derived from
deepseek-ai/DeepSeek-V4.1-Flash.
The conversion preserves the source affine 2-bit/group-64 values while stacking
routed experts for native batched execution. Conservative routing-saliency
pruning keeps 336 of 384 routed experts in every layer.
Qualification result
Hardware: Apple M3 Ultra, 60 GPU cores, 256 GiB unified memory, MLX 0.32.2.
| Measurement | Result |
|---|---|
| Tensor payload | 212,930,051,680 bytes (~199 GiB) |
| Strict resident load | 239.44 seconds |
| Peak MLX memory | 213.51 GB |
| Conservative decode | 7.31 tok/s |
| Best short-decode ceiling | 7.92 tok/s |
| Routed experts retained | 336 / 384 per layer |
| Mean retained routing mass | 99.9949% |
| Worst-layer retained routing mass | 99.8564% |
Decode throughput uses greedy generation and (N-1)/elapsed, excluding
prefill. The same greedy token chain was reproduced at evaluation intervals 4,
20, and 40. Correctly framed smoke output was The capital of France is Paris.
followed by EOS.
The artifact misses Rapid-MLX's 12 tok/s product floor by 34.0% and is therefore published for reproducibility and further optimization only. Uploading it does not mean that Rapid-MLX catalog, server, or Desktop support is available.
Initial evaluation results
These are the initial qualification checks completed before publication. They are intentionally reported as narrow pass/fail evidence, not as a general model quality score.
| Check | Result | Evidence |
|---|---|---|
| One-sentence factual chat | Pass | Prompt: What is the capital of France? Answer in one short sentence. Output: The capital of France is Paris. then EOS. |
| Greedy stability | Pass | Evaluation intervals 4, 20, and 40 produced the identical token sequence. |
| Prefill/decode numerical parity | Pass | Tiny-config prefill, decode, and chunked-prefill relative differences were at most 1.4e-6. |
| Next-token parity | Pass | 100% argmax agreement across the tiny-config comparison battery. |
| Quantized tensor fidelity | Pass | Quantize/dequantize cases were bit-exact; strict float and quantized layouts round-tripped. |
| Full artifact load | Pass | All 41 shards strict-loaded with no missing or unexpected text-model parameters. |
| Vision, tool use, long context | Not evaluated | These capabilities are not claimed by this artifact. |
The factual chat case used the required model framing:
<｜begin▁of▁sentence｜><｜User｜>What is the capital of France? Answer in one short sentence.<｜Assistant｜></think>
Raw prompts without these role markers produced invalid diagnostics and are not counted as evaluation results.
Calibration and validation
- REAP calibration used 2,048 real causal tokens split into two disjoint halves.
- Split-half expert keep-set overlap was 92.1354%.
- Tiny-config parity checks passed with relative differences no greater than 1.4e-6 and 100% argmax agreement.
- Quantize/dequantize cases were bit-exact; strict float/quantized layouts and streaming output matched.
- The 41-shard artifact strict-loaded without missing or unexpected text-model parameters.
These checks do not establish broad quality, tool-use, long-context, or safety parity with the original model. Two-bit quantization and expert pruning may reduce quality. Long context has not been qualified.
Experimental usage
Clone the exact qualification implementation and download the model through the standard Hugging Face cache:
git clone --branch v0.14.1 https://github.com/raullenchai/Rapid-MLX.git
cd Rapid-MLX
hf download rapid-mlx/DeepSeek-V4.1-Flash-REAP-2bit-MLX
python scripts/qualify_deepseek_v41_native.py \
~/.cache/huggingface/hub/models--rapid-mlx--DeepSeek-V4.1-Flash-REAP-2bit-MLX/snapshots/<revision> \
--tokens 16 --eval-intervals 4
Allow substantial memory headroom for macOS and context growth. This artifact is not suitable for machines below 256 GiB unified memory.
Background
The runtime is a native MLX implementation of the V4.1 architecture — attention,
MoE, engram, hyper-connections, indexer, sparse attention and a chunked-prefill
streaming generator — written for this qualification and shipped in Rapid-MLX
0.14.1 under scripts/deepseek_v41_native/.
Why the artifact is published even though it misses the product floor, and what the numbers mean in practice, is written up here: DeepSeek V4.1 on a Mac, and Why Your LLM API Router Is Reading Your Prompts.
For models that are product-qualified and run on an ordinary Mac, see rapidmlx.com.
License and attribution
The MIT license from the original release is included in LICENSE. Model
design and training belong to DeepSeek and its contributors. The source 2-bit
conversion was published by Vontra; this independent REAP repack was produced
by the Rapid-MLX project. See NOTICE for the pinned source revision.