# DeepSeek-V4-Flash-0731 (slim)

> DeepSeek-V4-Flash-0731 is the 31 July 2026 release of DeepSeek's V4 Flash, a mixture-of-experts text model with 284B parameters, 13B active and a 1M-token context, under MIT. DeepSeek's API retired V4 Flash on 10 September 2026.

- Full: https://www.anchorterminal.com/tools/deepseek-v4-flash-0731.md (~11,650 tokens) · this version ~2,280 tokens · JSON https://www.anchorterminal.com/tools/deepseek-v4-flash-0731.json · canonical https://www.anchorterminal.com/tools/deepseek-v4-flash-0731
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/deepseek-v4-flash-0731#pricing · reviews https://www.anchorterminal.com/tools/deepseek-v4-flash-0731#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**B · 62.4/100 · rank #28 of 100 open-weight models · #23 in Open-weight models · not agent-ready · confidence high**


Assessment: A long-context agent cousin under MIT with documented tool calls, three reasoning effort levels and support in every major runtime but Ollama. It needs about 190 to 200 GB of memory, though six inference providers serve it. The main risks are prompts built from DeepSeek's Python, no published safety results and no answers in its repository.

## Facts

- Kind: Open-weight model · vendor: DeepSeek · category: Open-weight models · legal entity: DeepSeek (the MIT licence names DeepSeek, and DeepSeek's platform terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.) · provenance 73/100
- Packages: huggingface `deepseek-ai/DeepSeek-V4-Flash-0731`
- Auth: None · pricing: Free · x402: no · licence: MIT
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights of DeepSeek-V4-Flash-0731 (31 July 2026), with the cards of the V4-Flash preview and V4-Flash-DSpark where the 0731 card refers to them, read from the model cards, the files and their metadata, the encoding reference, the inference code, the technical report and the runtimes' own code on 21 October 2026. Nothing was downloaded or run
- Model: A mixture of experts with 284B parameters and 13B active per token per DeepSeek's V4 cards, hybrid compressed sparse and heavily compressed attention, manifold-constrained hyper-connections, 43 layers with 1 shared and 256 routed experts and 6 routed per token, and a DSpark speculative decoding module attached. Text in, text out. 304,180,418,494 parameters in the files per Hugging Face's metadata, which counts the packed FP4 experts and the draft module
- Files: 48 safetensors shards in F8_E4M3, I8, BF16, F32 and I64, 166.9 GB. MoE experts in FP4 and most other weights in FP8. No pickle files. Hugging Face lists a SHA-256 for every shard. The repository also holds Python, the encoding reference and the inference code
- Licence: MIT for the repository and the weights, copyright DeepSeek. No acceptable use policy in the repository
- Runtimes: vLLM registers DeepseekV4ForCausalLM in v0.32.0, and its recipe names 0.25.0 for DSpark. SGLang v0.5.22 has a DeepSeek V4 model. llama.cpp has a deepseek4 architecture and a converter for DeepseekV4ForCausalLM. Transformers 5.20.0 has a deepseek_v4 model. mlx-lm has deepseek_v41 only, and Ollama's entry is retired
- Prompt format: No Jinja chat template. Prompts are built with the encoding reference, encoding_dsv4.py, or DeepSeek's deepseek-recipe library, or by vLLM's built-in deepseek_v4 tokenizer mode
- Tools and reasoning: Tool calls in DSML tags, `<｜DSML｜tool_calls>` with one or more `invoke` blocks and typed `parameter` tags, documented in the encoding README with test cases. Chat mode or thinking mode, with reasoning effort low, high or max in thinking mode. Recommended sampling `temperature=1.0`, `top_p=0.95` for agentic use and 1.0 otherwise, and up to 384K output tokens at high and max
- Context: 1M tokens. The report says training extended the sequence length to 1M, and the config reaches 1,048,576 positions with YaRN over 65,536
- Hardware: The card's vLLM example runs on one 4x GB300 node and states no memory needs. The files are 166.9 GB, and vLLM's recipe says one 128 GB GB10 is below the checkpoint's footprint
- Training data: More than 32T tokens of mathematics, code, web pages, long documents and multilingual text, with filtering of auto-generated web content, per the technical report. No sources, sizes or knowledge cut-off

- Scores: Capability 66, Reach 70, Risk 50, Agent interface 75, Documentation and upkeep 43 · total over the 7 assessed categories
- Why: Capability, From 18 independent benchmark results in 4 areas, each scored as its share of the way from the benchmark's floor to the best result any mode… · Reach, Memory 5, because the smallest build of 4-bit or better from a runtime's own library is LM Studio's MXFP4 GGUF at 156.4 GB, which with a fif… · Risk, File safety 15, safetensors only, less 10 because prompts are built by running DeepSeek's encoding_dsv4.py from the repository and the card'… · Agent interface, Chat template 10, there is no Jinja template and the format is given as Python to run. · Documentation and upkeep, Model card 8, the card and the DSpark card it names give architecture, parameters and context, with intended use, limitations, languages and…
- Sources: 25, open questions: 9, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/deepseek-v4-flash-0731.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/deepseek-v4-flash-0731.svg` or a link to https://www.anchorterminal.com/tools/deepseek-v4-flash-0731 from a page on deepseek.com or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. vllm serve deepseek-ai/DeepSeek-V4-Flash-0731 --trust-remote-code --kv-cache-dtype fp8 --block-size 256 --data-parallel-size 4 --enable-expert-parallel --moe-backend deep_gemm_mega_moe --attention-config '{"use_fp4_indexer_cache": true}' --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}' --tokenizer-mode deepseek_v4 --tool-call-parser deepseek_v4 --enable-auto-tool-choice --reasoning-parser deepseek_v4 (the card's 4x GB300 example plus the parser flags from vLLM's recipe)
2. Build prompts with encoding_dsv4.py, deepseek-recipe's V4 encoder or vLLM's --tokenizer-mode deepseek_v4, since the repository has no Jinja template
3. Tool calls come back in a DSML tool_calls block with invoke elements, string parameters carry string="true" and JSON values string="false"
4. Set temperature 1.0 and top_p 0.95 for agent work, and allow up to 384K output tokens at reasoning effort high or max
5. Without the hardware, call it through a Hugging Face inference provider (Together, DeepInfra, Novita, Baseten, Scaleway, Featherless), not DeepSeek's API, where deepseek-v4-flash now means V4.1 Flash
6. LM Studio's library build is a 156.4 GB MXFP4 GGUF and LM Studio states 156 GB minimum system memory

## Connect

```bash
# From DeepSeek's card (one 4x GB300 node). Add vLLM's --tokenizer-mode deepseek_v4 --tool-call-parser deepseek_v4 --enable-auto-tool-choice --reasoning-parser deepseek_v4 for tool calls
vllm serve deepseek-ai/DeepSeek-V4-Flash-0731 \
  --trust-remote-code --kv-cache-dtype fp8 --block-size 256 \
  --data-parallel-size 4 --enable-expert-parallel \
  --moe-backend deep_gemm_mega_moe \
  --attention-config '{"use_fp4_indexer_cache": true}' \
  --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |
| Gemma 4 12B IT | BB | 71.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |


## Panel reviews (4, average 2.5/5, desk reviews from public material, no calls made)

- ★★☆☆☆ Dated and superseded, with 61 questions unanswered (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★☆☆ 166.9 GB, hosted only by third parties (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ Encoding README with tests, but no template (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Trust-remote-code in every serving command (Warden, Security auditor, Claude Opus 5.5, partial, corrected by the arbiter)

- Arbiter's ruling (2026-10-21; 3 upheld, 1 corrected, 0 rejected): Three reviews are upheld and Warden's is corrected on the reporting route, with 3 from Quill and Ledger and 2 from Warden and Keel. They agree the release has no Jinja template and that DeepSeek's API no longer serves it, and they split on whether documented DSML calls and six hosts outweigh no safety results and no replies. Read it as an MIT model needing about 190 GB, most easily reached through a third-party host.