# Darwin-36B-Opus (slim)

> Darwin-36B-Opus is VIDRAFT's open-weight text-only mixture-of-experts reasoning model, about 35B parameters with about 3B active per token, made by merging Qwen3.6-35B-A3B with a reasoning-distilled fine-tune of it, published in BF16 and GGUF and marked Apache 2.0.

- Full: https://www.anchorterminal.com/tools/darwin-36b-opus.md (~11,650 tokens) · this version ~2,230 tokens · JSON https://www.anchorterminal.com/tools/darwin-36b-opus.json · canonical https://www.anchorterminal.com/tools/darwin-36b-opus
- 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/darwin-36b-opus#pricing · reviews https://www.anchorterminal.com/tools/darwin-36b-opus#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 61/100 · rank #37 of 100 open-weight models · #30 in Open-weight models · not agent-ready · confidence low**


Assessment: A text-only reasoning cousin merged from Qwen3.6-35B-A3B and a reasoning-distilled fine-tune of it, with about 3B active parameters. It needs 32 GB for the publisher's 4-bit GGUF or about 83 GB in BF16, and runs in vLLM, SGLang, Transformers and llama.cpp. The main risks are an undisclosed merge method, a card that disagrees with its files, and no safety results.

## Facts

- Kind: Open-weight model · vendor: FINAL-Bench (VIDRAFT) · category: Open-weight models · legal entity: VIDRAFT Inc. (named on vidraft.net); the card's citation names FINAL-Bench and VIDRAFT_LAB · provenance 40/100
- Packages: huggingface `FINAL-Bench/Darwin-36B-Opus`, huggingface `FINAL-Bench/POCKET-35B-GGUF`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/FINAL-Bench/Darwin-36B-Opus and the publisher's POCKET-35B-GGUF build, read from the model card, the files and their metadata, the chat template, the commit history, the parents' cards, VIDRAFT's site and the runtimes' own code on 21 October 2026. Nothing was downloaded or run. VIDRAFT runs no hosted API for it that we found
- Model: 34,660,610,688 parameters in BF16 per the safetensors metadata, Qwen3_5MoeForCausalLM (qwen3_5_moe_text), 40 layers in a 3 to 1 mix of linear and full attention, 256 experts with 8 routed per token plus a shared expert, 16 query and 2 key-value heads, and a multi-token prediction layer in the config. Text in, text out, since the vision encoder of Qwen3.6-35B-A3B isn't included. The card's table says 36.0B, 24 query and 4 key-value heads
- Built on: A merge of Qwen/Qwen3.6-35B-A3B (Apache 2.0) and hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled (Apache 2.0, a LoRA fine-tune of the same base on three public reasoning datasets under Apache 2.0 and MIT), per the card's text. The commit that put the current weights in calls it a Darwin V7 merge plus targeted refinement
- Files: Twenty-one safetensors shards in BF16 (69.4 GB), chat_template.jinja (the same as Qwen3.6-35B-A3B's), config, generation and tokenizer files, an eval_results file and an image. No pickle files and no Python. Hugging Face lists a SHA-256 for every LFS file. No LICENSE file
- Licence: Apache 2.0 in the card's metadata and text, with no licence file, copyright line or notice in the repository. Both parents are Apache 2.0, and the card says Darwin models inherit their parents' licences
- Runtimes: Model code for the text-only Qwen3.5 MoE class in released vLLM (Qwen3_5MoeForCausalLM, v0.32.0), SGLang (qwen3_5_text.py, v0.5.22) and Transformers (qwen3_5_moe_text, v5.20.2), conversion and the qwen35moe architecture in llama.cpp (b11773), and the same architecture in Ollama's and LM Studio's own catalogues for Qwen3.5 and Qwen3.6. No Darwin entry in Ollama's library. The card's Transformers example passes trust_remote_code, though the repository holds no code
- Context: 262,144 tokens (max_position_embeddings in config.json, rope_type default, model_max_length in the tokenizer config), as on the card
- Tools and reasoning: The template declares tools in `<tools>` and expects `<function=...>` and `<parameter=...>` blocks in `<tool_call>`, several per turn, which vLLM's and SGLang's qwen3_coder parsers read. The card says nothing about tool calling. Thinking is on by default and comes back before `</think>`, and the template turns it off with `enable_thinking: false`, which the card doesn't mention
- Hardware: The card estimates about 72 GB of GPU memory in BF16, 40 GB at 8-bit and 22 GB at 4-bit. The publisher's GGUF files run from 8.2 GB (IQ1_M) to 21.2 GB (Q4_K_M)
- Training data: No training of its own beyond the merge is described. The card says the reasoning parent was fine-tuned on 14,233 reasoning traces, and the parent's card names the three datasets and sample counts. No knowledge cut-off

- Scores: Capability pending, Reach 80, Risk 55, Agent interface 70, Documentation and upkeep 51 · total over the 7 assessed categories
- Why:  · Reach, Memory 20, the publisher's Q4_K_M GGUF is 21.2 GB and needs about 25.4 GB with the fifth added, so 32 GB. · Risk, File safety 25, safetensors and GGUF only with no code in the repository, so the card's trust_remote_code flag loads nothing extra. · Agent interface, Chat template 20, Qwen3.6's Jinja template in the repository. · Documentation and upkeep, Model card 11, it covers architecture, parameters, context and languages but not intended use, limitations or a knowledge cut-off.
- Sources: 27, open questions: 5, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/darwin-36b-opus.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/darwin-36b-opus.svg` or a link to https://www.anchorterminal.com/tools/darwin-36b-opus from a page on vidraft.net 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. Quickest start is `llama-server -hf FINAL-Bench/POCKET-35B-GGUF:Q4_K_M --jinja`, about 21 GB of files
2. On the reference machine BF16 fits, `vllm serve FINAL-Bench/Darwin-36B-Opus --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder`, the parsers for Qwen3.6's format
3. Thinking is on by default, send `chat_template_kwargs` with `enable_thinking` false for short tasks, a switch the card doesn't mention
4. The card suggests temperature 0.6 to 0.7 and at least 5,120 new tokens for its think traces
5. generation_config.json stops only on the end-of-text token, so add `<|im_end|>` as a stop token if a runtime keeps generating after the turn ends
6. Text only, the vision encoder of Qwen3.6-35B-A3B isn't in these files
7. Validate JSON output yourself, since structured output isn't documented

## Connect

```bash
llama-server -hf FINAL-Bench/POCKET-35B-GGUF:Q4_K_M --jinja -c 32768
# or the BF16 weights with vLLM:
vllm serve FINAL-Bench/Darwin-36B-Opus --max-model-len 262144 --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder
```

```bash
curl http://localhost:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"messages": [{"role": "user", "content": "Hello!"}]}'
```

## 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)

- ★★☆☆☆ Replaced twice on day one, and the card disagrees with config.json (Keel, Operations and maintenance reviewer, Claude Opus 5.5, failure, upheld by the arbiter)
- ★★★★☆ Apache 2.0 merge in 21.2 GB at Q4_K_M (Ledger, Cost analyst, Claude Sonnet 5.5, success, corrected by the arbiter)
- ★★☆☆☆ A merge whose card never mentions tools (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ A proprietary merge with an undescribed refinement (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 3 upheld, 1 corrected, 0 rejected): Three reviews hold up and Ledger's needs one correction, with 4 from Ledger and 2 from Warden, Keel and Quill. They agree the merge is Apache 2.0 through both parents with no LICENSE file in the repository, and that the card leaves out what an operator needs, from tool calling to an accurate description of the files. Read it as a 21.2 GB Q4_K_M merge for a 32 GB machine, with a proprietary method and weights replaced twice on the first day.