DeepSeek-V4.1-Flash-cmf

DeepSeek-V4.1-Flash → CMF — one file, one Rust runtime

DeepSeek-V4.1-Flash is converted to the CMF container: one memory-mapped weight file, served by a Rust binary without a Python ML framework. The runtime supports text and images, dynamic expert staging on Vulkan, and CPU fallback.

The model downloads as 59 verified parts and reconstructs into one file. Use the matching runtime supplied here; stock cortiq 0.6.6 predates this V4.1 integration. Download · Run · Measured results

This conversion derives from DeepSeek-V4.1-Flash, revision dba1be0a40aa45a94ad051997016db3960a90277.

Q4TP applies to ordinary text and expert matrix weights. Sensitive attention/control tensors retain unquantized higher precision. Vision and aligner matrices and controls remain unquantized F16, preserving the source BF16 precision class without introducing vision quantization loss. The two Engram lookup tables and their scale bytes remain lossless in their original FP8 E4M3 / E8M0 representation. They account for 202,758,032,400 bytes, so the complete artifact is a mixed-precision model. The CMF metadata retains the original nested text, vision and quantization configuration.

The accepted CMF file is dsv41-q4tp.cmf, exactly 498,986,143,769 bytes, with SHA-256 a681cafb28bea46271d1e28df5e532288ef545661541762ed06351dd1cf1cd0c.

Download

parts-q4tp/part_* concatenate into a single CMF file. The reconstruction manifest records every part's offset, byte size and SHA-256, plus the SHA-256 of the complete model.

Get the parallel downloader and Linux runtime:

curl -L https://huggingface.co/infosave/DeepSeek-V4.1-Flash-cmf/resolve/main/download_q4tp.py -o download_q4tp.py
curl -L https://huggingface.co/infosave/DeepSeek-V4.1-Flash-cmf/resolve/main/cortiq-linux-x86_64 -o cortiq-linux-x86_64

Use the downloader on Linux or macOS (Python 3.10+ is only needed to fetch and reconstruct the file):

python3 download_q4tp.py --workers 8 --output dsv41-q4tp.cmf

The script resolves one immutable Hub revision, downloads parts concurrently directly into the destination file, and checks both the parts and final model. It needs no second full copy of the weights. Run the same command after an interruption to reuse verified ranges. Keep the .partial.parts directory until the download completes.

Run

Download cortiq-linux-x86_64 and make it executable. The Linux binary, complete source archive, patch against public 0.6.6, dependencies, and build instructions are supplied together; see BUILD.md.

The tested host is one RTX PRO 6000 Blackwell with 96 GB VRAM, approximately 234 GiB available system RAM, and a 13.6-core CPU quota. The 498.99 GB model is memory-mapped from local disk and experts are staged dynamically. It does not fit wholly in VRAM. Allow at least 500 GB for the reconstructed model and additional space for the runtime and filesystem; fast local NVMe is preferable to a network volume.

Tested 96-GB profile (88 GiB configured GPU budget):

chmod +x cortiq-linux-x86_64
export XDG_RUNTIME_DIR=/tmp WGPU_BACKEND=vulkan CMF_GPU=1
export CMF_GPU_VRAM_MB=90112 CMF_THREADS=12 RAYON_NUM_THREADS=12
export CMF_DSV41_DYNAMIC_MOE=1 CMF_DSV41_GLOBAL_SEGMENTS=16
export CMF_DSV41_POOL_PCT=85 CMF_DSV41_FETCH_MAX=6
export CMF_DSV41_FETCH_MIN_SEEN=1 CMF_DSV41_GPU_ATTN=1
export CMF_DSV41_FUSED_Q=1 CMF_SERVE_SLOTS=1
./cortiq-linux-x86_64 serve ./dsv41-q4tp.cmf --host 127.0.0.1 --port 8080 --o1 off

The server exposes an OpenAI-compatible /v1 API. For CPU inference set CMF_GPU=0; the weight file and its disk requirement remain the same. The 16-segment profile is specific to the tested large GPU. Smaller budgets use 8 segments; a 16-GiB configured-budget check is reported separately from physical hardware testing.

Example request with the same non-thinking mode used for the measurements:

curl http://127.0.0.1:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"deepseek_v41-cortiq","messages":[{"role":"user","content":"Explain an LRU cache briefly."}],"temperature":0,"max_tokens":64,"enable_thinking":false}'

Performance

The selected fused-Q profile was measured with a single 64-token response, one serving slot, temperature zero and thinking disabled:

The stream's first event contains 22 tokens; 42 tokens arrive afterward. The rate above divides those 42 tokens by their arrival span, verified with the checkpoint tokenizer. It is not a whole-request rate or an engine-only decode timer. A first request means a newly started server, not a guarantee that the operating-system page cache is cold. These results do not reach 20 tokens/s and are not a universal throughput guarantee.

In the same-binary comparison, disabling fused Q produced 11.09 tokens/s on the repeated request; the 1.5% difference is small. The larger accepted improvement came from reusing the GPU attention/output frame: a prior matched comparison measured 11.75 versus 4.31 tokens/s with that frame enabled versus CPU attention. That comparison already used dynamic GPU MoE in both arms, so it is not a CPU-only comparison.

The implementation also reuses the shared dynamic expert pool, supports 16 pool segments for this 96-GB profile, and avoids a discarded host copy when the optional RAM tier is disabled. A larger 95% expert-pool budget and parallel refill-copy experiment did not produce useful measured gains and were not included in the recommended profile.

KV cache

The global cache stores main attention rows in E2M1 with E4M3 scales per 16 values, and index rows in E2M1 with E8M0 scales per 32 values. The four CSA2 source banks use 890 bytes per token in total, before fixed metadata and the bounded sliding windows. Actual storage at 4096 tokens is 3,645,440 bytes. Selected main rows are decoded once per layer into a bounded scratch buffer shared by the attention heads; the complete history is never unpacked for an attention call.

This reduces global-cache storage from the earlier f32 representation's 6400 bytes per token while preserving its decoded values. It does not reduce the 498.99 GB weight-file size. The official production engine also uses DSpark speculative decoding; this runtime retains those weights but does not yet execute DSpark or its batched verifier.

The official technical report describes additional production kernel fusion and speculative execution. Those production throughput claims do not transfer automatically to this portable CPU/Vulkan runtime and its weight-offloading hardware setup.

Validation

The accepted artifact gate covers all 48,498 retained tensors against 96,085 pinned source-index entries, including 40 layers, 384 routed experts per layer, two Engram layers, the full expert/mHC inventory, and three DSpark entries. CMF bounds, metadata, canonical coverage, duplicate detection, and native Engram U8 geometry passed. All MTP tensors are retained; retaining these tensors does not imply that speculative decoding is enabled. The artifact receipt is artifact-summary.json.

All 59 uploaded weight parts and the reconstruction manifest were checked against the accepted artifact at Hub revision dbabe4108fb2020630202b788f52590fd044fdfd.

The Engram component check covers the complete token mapping, hash parameters, text/image-break/padding cases, contiguous and chunked lookup, and actual selected FP8 rows from both layers. All 577,536 gathered BF16 values match the source exactly. The maximum residual-output difference is 0.015625; the mean absolute difference is at most 2.29e-7. These are component measurements, not full-model logit parity.

The final fused-Q GPU profile passed the semantic review of eight bounded cases: arithmetic, JSON, Russian, Python code, a counting problem, Chinese, conversation context, and an image. Seven of eight strict automated checks passed. Arithmetic correctly returned 3945 but included working despite an integer-only instruction. The generated parentheses-checking function passed 15 assertions. The image answer correctly counted 2 red squares, 2 blue circles and 1 green triangle, read CMF 41 and 8123, and took 85.92 seconds.

An earlier GPU profile without fused Q gave an incorrect total in the counting problem; the recommended fused-Q profile and CPU control both answered 18 correctly. The earlier failure is retained in the reports. These small checks are behavior samples, not a benchmark-suite score or full-checkpoint parity.

CPU multi-prompt checks, packed-cache component checks, and a two-prompt run with a configured 16-GiB GPU budget passed their documented semantic checks. The budget test ran on the physical 96-GB card; it does not establish speed or memory behavior on physical 16-GB hardware. The final fused-Q CPU fallback also matched the prior path exactly on the tiny fixture.

Runtime identities and measured text/image/resource results are recorded in release-manifest.json and the accompanying reports/ files.

Image example

Asked to return the counts, title and number as JSON, the model returned:

{"red_squares": 2, "blue_circles": 2, "green_triangles": 1, "title": "CMF 41", "number": 8123}

The accepted vision component comparison is component-level evidence only: the F16 Vulkan run produced finite output with cosine 0.999695 against the official component oracle. OpenRouter responses are a behavioral comparison only; they do not establish numerical parity with the original checkpoint. The available eight-case comparison was semantically correct in 8/8 cases and passed the strict automated format checks in 7/8 cases; the arithmetic case included an intermediate line despite requesting only the integer.

License

The source model uses the MIT license. CMF and the cortiq runtime are available from infosave2007/cmf.