# DeepSeek-V3.2 (slim)

> DeepSeek-V3.2 is DeepSeek's open-weight mixture-of-experts reasoning model with DeepSeek Sparse Attention, trained for tool use with a 128K-token context, published in FP8 under MIT. DeepSeek's API stopped serving it on 24 April 2026, moving to V4 Flash.

- Full: https://www.anchorterminal.com/tools/deepseek-v3-2.md (~11,150 tokens) · this version ~2,280 tokens · JSON https://www.anchorterminal.com/tools/deepseek-v3-2.json · canonical https://www.anchorterminal.com/tools/deepseek-v3-2
- 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-v3-2#pricing · reviews https://www.anchorterminal.com/tools/deepseek-v3-2#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 58.6/100 · rank #65 of 100 open-weight models · #54 in Open-weight models · not agent-ready · confidence high**


Assessment: DeepSeek-V3.2 is an MIT-licensed reasoning model trained for parallel tool calls, with thinking chosen per request, for long agent work. It needs about 690 GB of FP8 files, eight 80 GB GPUs in vLLM's recipe and prompts built from DeepSeek's Python. The main risk is neglect, with no file changes since December 2025 and no safety evaluation.

## 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 60/100
- Packages: huggingface `deepseek-ai/DeepSeek-V3.2`
- Auth: None · pricing: Free · x402: no · licence: MIT
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/deepseek-ai/DeepSeek-V3.2, read from the model card, the technical report in the repository, the files and their metadata, the encoding reference, the inference code, DeepSeek's DeepSeek-V3.2-Exp GitHub repository (which the card points to for running it) and the runtimes' own code on 21 October 2026. Nothing was downloaded or run. DeepSeek-V3.2-Speciale, a separate repository with the same architecture, isn't graded here
- Model: A mixture of experts with DeepSeek Sparse Attention (a lightning indexer picking 2,048 key-value tokens per query) on multi-head latent attention, 61 layers, 1 shared and 256 routed experts with 8 routed per token, and one multi-token prediction layer. Text in, text out. The card states no parameter count; the inference config is named 671B and Hugging Face's metadata counts 685,355,329,792 parameters in the files. Continued from DeepSeek-V3.1-Terminus through DeepSeek-V3.2-Exp-Base
- Files: 163 safetensors shards in F8_E4M3 with BF16 and F32 for the rest, 689.5 GB stored. No pickle files. Hugging Face lists a SHA-256 for every shard. The repository also holds the encoding reference and DeepSeek's inference demo in Python
- Licence: MIT for the repository and the weights, copyright DeepSeek. No acceptable use policy in the repository
- Runtimes: Model code in vLLM (DeepseekV32ForCausalLM, with a deepseek_v32 tool parser and a recipe), SGLang (DeepseekV32ForCausalLM, a deepseekv32 tool-call detector and a cookbook page), llama.cpp (the deepseek32 architecture and its converter), mlx-lm (deepseek_v32) and Transformers (deepseek_v32, present in the released 5.15.0 and 5.20.0). Ollama's library entry was retired on 15 July 2026 with no models pushed. DeepSeek's own inference folder is a demo for torchrun across several GPUs
- Prompt format: No Jinja chat template, as the card says. Prompts are built with DeepSeek's Python encoding reference, encoding/encoding_dsv32.py, with test inputs and outputs. llama.cpp, SGLang and mlx-lm ship templates of their own
- Tools and reasoning: Tools are declared as JSON Schema in the system prompt and calls come back in DSML tags, a `function_calls` block holding one or more `invoke` elements with typed `parameter` elements. A `response_format` schema can be added to the system prompt. Thinking mode or chat mode is chosen per request, and earlier reasoning is dropped by default. A `developer` role is for search agents only. Recommended sampling `temperature=1.0`, `top_p=0.95`
- Context: 128K tokens per the technical report, which continued pre-training on 128K-token sequences. The config sets 163,840 positions through YaRN (factor 40 over 4,096) and the tokeniser config 131,072
- Hardware: The card states no memory needs. The files are 689.5 GB, DeepSeek's demo shards them across GPUs with torchrun, and vLLM's recipe asks for at least 8 H100 or H200 80 GB GPUs. It doesn't fit in 128 GB or 512 GB
- Training data: The report describes continued pre-training of 2.1B tokens for the indexer warm-up and 943.7B tokens of 128K-token sequences aligned with V3.1-Terminus's long-context data, then reinforcement learning on synthesised agent tasks. No sources, sizes by source or knowledge cut-off

- Scores: Capability 69, Reach 60, Risk 50, Agent interface 60, Documentation and upkeep 34 · total over the 7 assessed categories
- Why: Capability, From 19 independent benchmark results in 5 areas, each scored as its share of the way from the benchmark's floor to the best result any mode… · Reach, Memory earns nothing, because the only build from DeepSeek or a runtime's library is the 689.5 GB FP8 release, which needs about 827 GB with… · Risk, File safety earns 15, safetensors only but 10 off because prompts are built by running DeepSeek's Python from the repository. · Agent interface, The chat template earns 10, as the format exists only as Python code to run. · Documentation and upkeep, The model card earns 14 when read with DeepSeek's own V3.2 model card PDF, which adds intended use and a parameter count, leaving languages…
- Sources: 27, open questions: 6, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/deepseek-v3-2.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/deepseek-v3-2.svg` or a link to https://www.anchorterminal.com/tools/deepseek-v3-2 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. Start it with vLLM's recipe, vllm serve deepseek-ai/DeepSeek-V3.2 --tensor-parallel-size 8 --tokenizer-mode deepseek_v32 --enable-auto-tool-choice --tool-call-parser deepseek_v32 --reasoning-parser deepseek_v3
2. It doesn't fit the 107.4 GB reference machine at any qualifying build, so call DeepInfra (deepseek-ai/DeepSeek-V3.2) or SambaNova (DeepSeek-V3.2) instead
3. Outside vLLM, build prompts with encoding/encoding_dsv32.py and wrap parse_message_from_completion_text in error handling, since DeepSeek says it handles well-formed output only
4. Pick thinking or chat with thinking_mode, and read the reasoning from reasoning_content
5. Send developer messages only for search-agent prompts, the only use DeepSeek gives that role
6. Use temperature 1.0 and top_p 0.95 and keep prompts within 128K tokens

## Connect

```bash
# vLLM's own recipe for DeepSeek-V3.2 (8 GPUs)
vllm serve deepseek-ai/DeepSeek-V3.2 --tensor-parallel-size 8 \
  --tokenizer-mode deepseek_v32 --enable-auto-tool-choice \
  --tool-call-parser deepseek_v32 --reasoning-parser deepseek_v3
# or DeepSeek's reference demo, from the inference folder
python convert.py --hf-ckpt-path "$HF_CKPT_PATH" --save-path "$SAVE_PATH" --n-experts 256 --model-parallel 8
torchrun --nproc-per-node 8 generate.py --ckpt-path "$SAVE_PATH" --config config_671B_v3.2.json --interactive
```

## 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 2.3/5, desk reviews from public material, no calls made)

- ★★☆☆☆ Seven commits on one day, and FlashMLA has moved on (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ 689.5 GB and no first-party API left (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ DSML in Python, with no runtime parser from DeepSeek (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Prompts built by DeepSeek's Python, reports to one inbox (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with 3 from Quill and 2 from Warden, Keel and Ledger. They agree DeepSeek has left the release alone since 1 December 2025, with no Jinja template, no safety results and no first-party API since 24 April 2026. Read it as an MIT model for teams with an 8-GPU server, or one to call through DeepInfra or SambaNova.