# Qwopus3.8-27B-Flash-V2 (slim)

> Qwopus3.8-27B-Flash-V2 is Jackrong's open-weight 27B fine-tune of Qwen3.8-27B, by way of his Qwopus3.8-27B-Flash, trained for shorter reasoning in agent loops, published in BF16 and GGUF with a multi-token prediction head under Apache 2.0.

- Full: https://www.anchorterminal.com/tools/qwopus3-8-27b-flash-v2.md (~10,550 tokens) · this version ~2,080 tokens · JSON https://www.anchorterminal.com/tools/qwopus3-8-27b-flash-v2.json · canonical https://www.anchorterminal.com/tools/qwopus3-8-27b-flash-v2
- 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/qwopus3-8-27b-flash-v2#pricing · reviews https://www.anchorterminal.com/tools/qwopus3-8-27b-flash-v2#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 61/100 · rank #46 of 100 open-weight models · #39 in Open-weight models · not agent-ready · confidence low**


Assessment: A Qwen3.8-27B cousin tuned to think less per turn in long agent loops, with an MTP head for speculative decoding in llama.cpp. It fits 24 GB with Jackrong's 4-bit GGUF. The risks are one person's experimental release with no safety work, training traces taken from Claude and GPT, and a tool format that differs from Qwen's.

## Facts

- Kind: Open-weight model · vendor: Jackrong · category: Open-weight models · legal entity: Jackrong, an individual Hugging Face account · provenance 40/100
- Packages: huggingface `Jackrong/Qwopus3.8-27B-Flash-V2`, huggingface `Jackrong/Qwopus3.8-27B-Flash-V2-GGUF`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0 (metadata only, no LICENSE file); built on Qwen3.8-27B under Apache 2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/Jackrong/Qwopus3.8-27B-Flash-V2 and Jackrong's GGUF repository, read from the model cards, the files and their metadata, the chat template, the base models' metadata and licence, and the runtimes' support for the qwen3_5 architecture on 21 October 2026. Nothing was downloaded or run
- Model: 27,781,427,952 parameters in BF16 per the Hub's safetensors count, identical to Qwen3.8-27B's, Qwen3_5ForConditionalGeneration (qwen3_5) with the vision encoder and one multi-token prediction layer. Text, images and video in, text out
- Built on: A further post-training run (new rewards and RL) on Jackrong/Qwopus3.8-27B-Flash, itself an SFT and reasoning-RL fine-tune of Qwen/Qwen3.8-27B. All three are Apache 2.0. Declared in the card's metadata (base_model Jackrong/Qwopus3.8-27B-Flash, relation finetune)
- Files: 18 safetensors shards (55.6 GB in all) with chat_template.jinja and processor config. The GGUF repository holds BF16, Q8_0, Q6_K, Q5_K_M, Q5_K_S, Q4_K_M, Q4_K_S, IQ4_XS, Q3_K_L, Q3_K_M, Q3_K_S and Q2_K files, each with the MTP head, and an F32 vision projector. No pickle files and no Python. No LICENSE file
- Licence: Apache 2.0 in the card's metadata, with no licence text in the repository. The base Qwen3.8-27B and the intermediate Qwopus3.8-27B-Flash are Apache 2.0 too. No use policy. The card says the training data includes traces derived from Claude and GPT, whose providers' terms aren't part of the model's licence
- Runtimes: The qwen3_5 architecture runs in released Transformers, vLLM, SGLang, llama.cpp (qwen35) and Ollama. The card's tests ran on llama.cpp with the bundled MTP head for self-speculative decoding. No runtime names it
- Context: 262,144 tokens in config.json and tokenizer_config.json, Qwen3.8-27B's native context. The card states none, and reports tests at 64K and 327,680
- Tools and reasoning: The template declares tools as JSON inside `<tools>` and asks for calls as `<tool_call>{"name": ..., "arguments": ...}</tool_call>`, several per turn, the Hermes shape, not Qwen3.8's. The card documents `enable_thinking` and `reasoning_effort` (xhigh, medium or low) as template arguments, and `preserve_thinking` keeps earlier reasoning
- Hardware: The card states no memory needs and reports tests on NVIDIA GPUs including one RTX 5090. Files are 55.6 GB in BF16 and 15.4 to 29.0 GB for Jackrong's 4-bit to 8-bit GGUFs
- Training data: About 1.5 million teacher examples cut to the best 10% by a judge ensemble (Qwen3.7-Max, GLM-5, GPT-OSS-120B, Gemma4-27B), plus agent trajectories and traces derived from Claude and GPT, then RL. The V2 recipe isn't disclosed. No sources, sizes or cut-off

- Scores: Capability pending, Reach 85, Risk 50, Agent interface 60, Documentation and upkeep 56 · total over the 7 assessed categories
- Why:  · Reach, Memory 25 of 30, Jackrong's IQ4_XS GGUF is 15.4 GB, about 18.5 GB with a fifth added, which fits 24 GB. · Risk, File safety 25, safetensors and GGUF only, nothing from the repository runs. · Agent interface, Chat template 20, a Jinja file in the repository. · Documentation and upkeep, Model card 11 of 20, missing architecture, context and a cut-off.
- Sources: 17, open questions: 4, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/qwopus3-8-27b-flash-v2.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/qwopus3-8-27b-flash-v2.svg` or a link to https://www.anchorterminal.com/tools/qwopus3-8-27b-flash-v2 from a page under huggingface.co/jackrong, 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 Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M --jinja`, about 17 GB, and add mmproj-F32.gguf for images
2. The card's tests used Q5_K_M with the bundled MTP head and `--spec-type draft-mtp` in llama.cpp
3. The template asks for tool calls as JSON inside tool_call tags, so use a Hermes-style parser (vLLM's hermes), not Qwen3.8's qwen3_coder
4. For coding the card recommends enable_thinking true, reasoning_effort xhigh and temperature 0.3, which you must pass yourself, since generation_config.json says 1.0
5. reasoning_effort takes xhigh (default), medium or low, and any other value makes the template raise an error
6. Earlier reasoning stays in the context unless preserve_thinking is false
7. Cap max_tokens, since a report says it loops more than V1

## Connect

```bash
llama-server -hf Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M --jinja -c 65536
# or, with vLLM:
vllm serve Jackrong/Qwopus3.8-27B-Flash-V2 --max-model-len 131072 --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser hermes
```

```bash
curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"messages": [{"role": "user", "content": "Hello!"}], "temperature": 0.3, "chat_template_kwargs": {"enable_thinking": true, "reasoning_effort": "xhigh"}}'
```

## 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)

- ★★★☆☆ One commit on 21 September, from a personal account (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★☆☆ 15.42 GB at IQ4_XS from a personal account (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Tool format changed from the base without a note (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ A personal account and a card that admits no safety testing (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 ratings from 2 to 3. They agree it is an individual's fine-tune of a fine-tune of Qwen3.8-27B, Apache 2.0 through the chain with no LICENSE file, and that the template switches tool calls to a Hermes-style format the card doesn't describe. Read it as a 24 GB agent fine-tune with publisher GGUFs, no safety testing by the card's own account and nobody independent benchmarking it.