Model · DeepSeek · Text
DeepSeek V4 Flash
DeepSeek's efficient tier of the V4 generation. Faster and cheaper than V4 Pro; the practical default for high-throughput agentic workloads, and MIT-licensed.
- Modality
- Text
- License
- MIT (Open)
- Parameter size
- 284B total, 13B activated
- Context window
- 1,048,576 tokens 1M-token context with up to 384k output.
- Released
- July 31, 2026
- Last verified
- September 8, 2026
- Runs locally
- Yes
Strengths
- MIT licence — no commercial restrictions on the weights
- Outperforms the V4 Pro preview on benchmarks despite a far smaller activated parameter count
- Large agentic gains over the preview — Terminal Bench 2.1 82.7 (from 61.8), DeepSWE 54.4 (from 7.3)
- Ships with a speculative-decoding module for faster inference
Weaknesses
- Trails Claude Opus 4.8 on repo-scale work (NL2Repo 54.2 vs 69.7)
- Slightly behind V4 Pro on pure knowledge tasks and the most complex agentic workflows
- 284B total parameters still needs multi-GPU serving — not a single-machine model
Try it
| Where | Type | Notes |
|---|---|---|
| Hugging Face | weights | MIT licence |
| DeepSeek Platform | hosted-api | API key required |
Used in solutions
Version history
- DeepSeek V4 Pro Aug 2026
- DeepSeek V4 Flash Jul 2026 Current
- DeepSeek V3 Dec 2024 Deprecated
Official sources
- Model card model
- Technical report paper
Change log
- — Corrected on three counts. Release date 2026-05-06 was a Hugging Face push timestamp, not a release — the official DeepSeek-V4-Flash-0731 shipped 2026-07-31 per DeepSeek's API changelog, superseding the preview this entry described. Licence corrected from 'DeepSeek License' to MIT, and the false 'restrictions on commercial use cases' weakness removed. Added parameter size and 1M context window.
- — Initial entry. Sighted on Hugging Face (lastModified 2026-05-06).
Esc