Views
No views yet
deepseek-ai/DeepSeek-V4-Flash-0731
(304B MoE, MIT license) with an extended reasoning budget, 1M-token context,
and DSpark speculative decoding — tuned on private, user-owned corpora only.| Field | Value |
|---|---|
| Base | deepseek-ai/DeepSeek-V4-Flash-0731 |
| Architecture | Mixture-of-Experts (MoE) + speculative decoding (DSpark) |
| Parameters | 304B total, fraction activated per token |
| Context window | 1,048,576 tokens (1M) |
| Max output | 384K tokens (high/max reasoning) |
| Reasoning effort | low / high / max |
| Precision | BF16 / FP16 / FP32 / FP8 (E4M3, E2M1) / INT8 |
| Quantizations | 90+ model tree (GGUF/safetensors) |
| License | MIT (derived) |
| Paper | arXiv:2606.19348 |
| Sovereign property | weights + fine-tunes stored locally; zero mandatory telemetry |
router.huggingface.co/v1)
is an option, not a dependency.1# vLLM (4×GB300 node) — DSpark speculative decoding enabled with one flag
2vllm serve deepseek-ai/DeepSeek-V4-Flash-0731 \
3 --trust-remote-code --kv-cache-dtype fp8 --block-size 256 \
4 --data-parallel-size 4 --enable-expert-parallel \
5 --moe-backend deep_gemm_mega_moe \
6 --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'
7
8# SGLang
9sglang serve --trust-remote-code \
10 --model-path deepseek-ai/DeepSeek-V4-Flash-0731 \
11 --tp 4 --moe-runner-backend flashinfer_mxfp4 \
12 --speculative-algorithm DSPARK --chunked-prefill-size 4096temperature=1.0, top_p=0.95;
otherwise top_p=1.0.encoding_dsv4.encode_messages(messages, thinking_mode="thinking", reasoning_effort="max") (see models/deepseek-v4-flash-0731/ docs and
api_examples/).