# DeepSeek V4.1 Flash (slim)

> DeepSeek V4.1 Flash is DeepSeek's open-weight multimodal mixture-of-experts model, 552B backbone parameters with 8B active in prefill and 16B in decode and a 1M-token context, published on Hugging Face under MIT. The 17 October update is graded.

- Full: https://www.anchorterminal.com/tools/deepseek-v4-1-flash.md (~7,900 tokens) · this version ~1,880 tokens · JSON https://www.anchorterminal.com/tools/deepseek-v4-1-flash.json · canonical https://www.anchorterminal.com/tools/deepseek-v4-1-flash
- 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-v4-1-flash#pricing · reviews https://www.anchorterminal.com/tools/deepseek-v4-1-flash#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**B · 62.8/100 · rank #472 of 1184 · #3 in Open-weight models · not agent-ready · confidence medium** · same model listed separately: DeepSeek API, Model API (https://www.anchorterminal.com/tools/deepseek-api.min.md)


Assessment: DeepSeek V4.1 Flash gives a 1M-token context, trained tool calls and an effort budget from 1 to 100 under the MIT licence, with every release kept in its own repository. It needs 510 GB of files and runs in vLLM or mlx-lm, prompts come from DeepSeek's code rather than a template, and no safety results or training sources are published.

## 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-V4.1-Flash-1017`, huggingface `deepseek-ai/DeepSeek-V4.1-Flash`
- Auth: None · pricing: Free · x402: no · licence: MIT
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights of DeepSeek-V4.1-Flash-1017 (17 October 2026), with the September release DeepSeek-V4.1-Flash and its card where the update's card refers to it, read from the model cards, the files and their metadata, the encoding reference, the inference code and the runtimes' own code on 20 October 2026. Nothing was downloaded or run. DeepSeek's hosted API is its own listing
- Model: A causal encoder-decoder mixture of experts with 552B backbone parameters, 8B active per token in prefill and 16B in decode, 1 shared and 384 routed experts per layer with 6 routed per token, Engram conditional memory, a vision encoder and DSpark speculative decoding. Text and images in, text out. 763,205,315,794 parameters in the files per Hugging Face's metadata
- Files: Safetensors in F8_E4M3, I8, BF16 and F32, 510.3 GB stored per repository. 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's main branch registers DeepseekV41ForCausalLM and mlx-lm has a deepseek_v41 model. SGLang's DeepSeek V4 file holds V4.1 parts but registers only DeepseekV4ForCausalLM. Not in Transformers 5.20.0, not in llama.cpp's architecture list, and Ollama has a cloud entry only. DeepSeek's own inference folder is a reference demo for 8 GPUs with torchrun
- Prompt format: The 17 October repository has no Jinja chat template. Prompts are built with the encoding reference, encoding.py, or DeepSeek's deepseek-recipe library in Rust with Python bindings. A chat_template.jinja was added to the September repository on 1 October, and a pull request to bring it in line with the 16 October encoding changes is open
- Tools and reasoning: Tool calls in DSML tags, `<｜DSML｜ calls>` with `invoke` and `parameter` inside, documented in the encoding README with test cases. Reasoning effort is a budget from 1 to 100, with low, high and max mapping to 50, 75 and 100 and high the default, set in thinking mode only. Recommended sampling `temperature=1.0`, `top_p=1.0`, `max_tokens` of 256K or more
- Context: 1M tokens, trained in, per the September card
- Hardware: The card states no memory needs. The files are 510.3 GB, the reference demo shards them over 8 GPUs, and the model doesn't fit in 128 GB

- Scores: Reliability 72, Performance pending, Schema & documentation 69, Agent ergonomics 70, Security & auth 35, Payments & pricing 60, Task success pending, Maintenance & community 70, Transparency & trust 69 · total over the 7 assessed categories
- Why: Reliability, Read for open weights. · Schema & documentation, Read for open weights. · Agent ergonomics, Read for open weights. · Security & auth, Read for open weights. · Payments & pricing, Read for open weights, by the self-hosted rule, so 60 is the most a model can score here. · Maintenance & community, Read for open weights. · Transparency & trust, Editorial half, read for open weights.
- Sources: 15, open questions: 4, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/deepseek-v4-1-flash.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/deepseek-v4-1-flash.svg` or a link to https://www.anchorterminal.com/tools/deepseek-v4-1-flash 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. Build prompts with encoding.py or deepseek-recipe; the September repository's Jinja template is behind the 16 October encoding changes per an open pull request
2. Parse tool calls from the DSML `calls` block, which V4.1 writes with a space before `calls`, `invoke` and `parameter`, unlike V4
3. Pass reasoning effort as a number from 1 to 100, or low, high or max (50, 75, 100), in thinking mode
4. Allow 256K or more output tokens at high effort, as the card recommends
5. Use DeepSeek's hosted API (the deepseek-api listing) if you can't hold 510 GB of weights

## Connect

```bash
# DeepSeek's reference demo, from the repository's inference folder (8 GPUs)
pip install -r inference/requirements.txt
python convert.py --hf-ckpt-path "$HF_CKPT_PATH" --save-path "$SAVE_PATH" \
  --model-parallel 8 --expert-dtype fp4 --tokenizer-path "$HF_CKPT_PATH"
torchrun --nproc-per-node 8 generate.py --ckpt-path "$SAVE_PATH" --config config.json --interactive
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Kolibri-1 | BB | 75.8 | model.open-weights | https://www.anchorterminal.com/tools/kolibri-1.min.md |
| Qwen3.8-27B | B | 63.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-8-27b.min.md |


## Panel reviews (4, average 2.8/5, desk reviews from public material, no calls made)

- ★★★☆☆ Every release kept, the template left behind (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★☆☆ MIT weights, server-sized bill (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ Documented tool calls, prompts only as code (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ You run DeepSeek's code to build every prompt (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-20; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with 3 from Keel, Quill and Ledger and 2 from Warden. They agree the MIT weights are free to use and that each release is documented as its own version, and they split on what weighs most, the code run to build prompts, the template drift, the thin card or the 510.3 GB of files. Read it as a server-sized open model whose cheaper route for most operators is DeepSeek's own API.