[!IMPORTANT] Generation is validated on one physical NVIDIA DGX Spark. Use the pinned Docker recipe below. Stock vLLM, stock ExLlamaV3, and a bare --tensor-parallel-size 1 command cannot load this TP4 rank-sliced artifact correctly.

This is a REAP-pruned, 3.0 bpw EXL3/Trellis build of deepseek-ai/DeepSeek-V4-Flash-0731, packaged for a single 128 GB DGX Spark through local-inference-lab/sparkinfer.

The supported runtime is public and reproducible:

Runtime repository: 0xSero/deepseek-v4-flash-0731-spark-sparkinfer
Pinned public image: ghcr.io/0xsero/deepseek-v4-flash-0731-spark-sparkinfer@sha256:2e077489a83a0360952828051fe7f7a32c1801e5ce8436d85f7267583d614ff4
Pinned model revision: 22f28d32b9b29b4352eaa380ff8c2c170b2847ab
API: OpenAI-compatible server on http://127.0.0.1:8000
Configured model limit: 262,144 tokens

Run it

1. Prepare the Spark

You need:

one NVIDIA DGX Spark with its GB10/SM121 GPU and 128 GB unified memory;
Linux ARM64 with current NVIDIA drivers;
Docker Engine, Docker Compose v2, and NVIDIA Container Toolkit;
at least 250 GB free disk for the image, downloaded TP4 checkpoint, losslessly coalesced TP1 checkpoint, compact draft, and caches. 300 GB is a safer allowance.

The model and GHCR image are public. No Hugging Face or GHCR login is required.

Confirm Docker can see the GPU:

docker run --rm --gpus all nvcr.io/nvidia/cuda:13.0.1-base-ubuntu24.04 nvidia-smi

2. Install and start with one command

bash

git clone https://github.com/0xSero/deepseek-v4-flash-0731-spark-sparkinfer.git && \
cd deepseek-v4-flash-0731-spark-sparkinfer && \
docker compose up -d

Do not replace the image digest or manually reconstruct the vLLM command. The Compose file pins the validated ARM64 runtime, model revision, SparkInfer patches, TP1 coalescer, K64 speculative draft, FP8 sparse-MLA compatibility cache, and CUDA-graph configuration.

3. Follow the first start

docker compose logs -f

The first start is intentionally long. It will:

pull the pinned ARM64 runtime image;
download this exact model revision (about 107 GB of published files);
losslessly coalesce its TP4 rank slices into a TP1 serving checkpoint;
verify the tensor manifest and SHA-256 checksums;
build and validate a separate 3.0 GB K64 DSpark speculative draft;
run the SparkInfer CUDA-graph self-test; and
load the model, capture FULL/PIECEWISE CUDA graphs, allocate the KV cache, and start port 8000.

Downloaded weights and compiled caches remain under ./data and ./cache. Restarting the container reuses valid artifacts instead of downloading or rebuilding them.

4. Wait for health

bash

until curl -fsS http://127.0.0.1:8000/health; do sleep 10; done
curl -fsS http://127.0.0.1:8000/v1/models

Healthy startup logs should show all of the following:

max_model_len=262144;
FULL_AND_PIECEWISE CUDA graphs, not eager execution;
fixed K5 DSpark verification with a six-row C1 graph;
at least 262,144 post-capture KV tokens; and
served model name deepseek-v4-flash-0731-spark.

5. Generate

bash

curl -sS http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "deepseek-v4-flash-0731-spark",
    "messages": [
      {"role": "user", "content": "Write a correct Python function that returns the first n Fibonacci numbers."}
    ],
    "temperature": 0,
    "max_completion_tokens": 512
  }'

Thinking is disabled in the measured fast profile. To request it for an individual call, pass the model's supported chat-template option explicitly; expect different latency and memory behavior.

Stop or restart

bash

docker compose stop
docker compose start

To remove only the running container while preserving downloaded model data and caches:

docker compose down

What was actually validated

The published Docker image was built from the public runtime repository and tested on one physical DGX Spark. The following acceptance surfaces passed:

complete model manifest and checksum verification;
all 216 target experts and target EXL3/Trellis tensors preserved;
lossless TP4-to-TP1 coalescing;
real semantic and code generation;
strict JSON-schema structured output;
exact beginning, middle, and end fact recovery at 500, 2,000, 5,000, 10,000, and 20,000 characters;
max_model_len=262144 with 264,867 post-capture KV tokens in the clean-image acceptance run;
FULL/PIECEWISE CUDA graphs with eager execution disabled; and
healthy serving with zero container restarts.

Measured C1 code decode across five 512-token trials was:

Metric	Result
Minimum	34.30 tok/s
Median	38.12 tok/s
Mean	39.49 tok/s

A same-runtime-path cold 252,047-token prefill run measured 1,055.45 prompt tok/s with zero cached prompt tokens. That prefill result predates the final published-image rerun, so it remains candidate evidence. The requested steady 35 tok/s minimum is also still an open optimization gate because one of the five clean-image decode trials measured 34.30 tok/s.

Full evidence and reproduction scripts are in the runtime repository's VALIDATION.md and results/.

What the formats and runtime do

REAP selected and compacted the routed experts using router-weighted activation observations. The target retains 216 of 256 experts per MoE scope. Tool-calling and agentic structured-output specialists were made mandatory before the remaining slots were filled by global saliency.
EXL3 is the low-bit weight format. It stores the retained routed experts near 3.0 bits per weight while carried DeepSeek FP8 tensors remain FP8.
Trellis allocates and reconstructs low-bit weights non-uniformly, keeping more precision where it matters instead of forcing every group to the same width. TR3 is the K3 Trellis tier used here.
SparkInfer provides the GB10/SM121 sparse-MLA and Trellis GPU paths used by the patched vLLM runtime.
DSpark/MTP supplies a separate speculative draft. The runtime derives a compact K64 draft from the preserved REAP rankings; it does not alter the K216 target checkpoint.

In short: REAP chooses the experts, EXL3/Trellis compresses and executes their weights, SparkInfer supplies the Spark-native kernels, and DSpark proposes tokens that the untouched target verifies.

Checkpoint provenance

Source EXL3: 0xSero/DeepSeek-V4-Flash-0731-EXL3-3.0bpw
Original EXL3 weight files: 124,867,114,600 bytes
This checkpoint: 106,816,685,560 bytes (99.48 GiB)
Removed: 18,050,429,040 bytes (18.05 GB)
Routed experts: 216 of 256 retained per MoE scope
Router top-k: 6, unchanged
Observation dataset: 0xSero/deepseek-v4-flash-reap-observations-v2
Observation snapshot: partial-v2-21289, 21,289 rows, all experts observed
Structured observations: 2,253 agentic-tool rows and 4,096 tool-calling rows
Keep-plan SHA-256: cc2bdaf189ebdaa98178e35abac4b1b336808c4f86b11e9b8eccde6cb3e44001

The ranking was transferred from the closely aligned prior DeepSeek V4 Flash revision. All 43 router matrices passed identity alignment; the weakest same-index cosine was 0.9460793734 and all three hash-routing tables were bit-identical.

Important implementation note

The validated single-Spark route uses a 584-byte padded FP8 sparse-MLA record under the nvfp4_ds_mla control path. It is not a true 432-byte NVFP4 KV record. The experimental 432-byte implementation passed its isolated kernel oracle but produced corrupted full-model text, so the Docker recipe disables it. The server-level generation and structured-output tests are the authoritative acceptance gates.

Credits

Thanks to DeepSeek for the base model, Cerebras Research for REAP, TurboDerp and contributors for ExLlamaV3/EXL3/Trellis, brandonmusic for the GLM-5.2 EXL3 TR3 3.0 bpw reference, Local Inference Lab for SparkInfer and the DGX Spark vLLM work, MiaAI-Lab for the DSpark reference, NVIDIA for the CUDA stack and DGX Spark, and the vLLM community.

Model use remains subject to the upstream model license and terms.