# Qwen3.8-Flash-Next (slim)

> Qwen3.8-Flash-Next is Alibaba's open-weight multimodal mixture-of-experts model and an experimental preview of the Qwen4 architecture, 125B parameters with 6B active plus a 51B n-gram embedding table, published on Hugging Face in BF16 and FP8.

- Full: https://www.anchorterminal.com/tools/qwen3-8-flash-next.md (~13,700 tokens) · this version ~2,480 tokens · JSON https://www.anchorterminal.com/tools/qwen3-8-flash-next.json · canonical https://www.anchorterminal.com/tools/qwen3-8-flash-next
- 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/qwen3-8-flash-next#pricing · reviews https://www.anchorterminal.com/tools/qwen3-8-flash-next#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 61.8/100 · rank #31 of 100 open-weight models · #25 in Open-weight models · not agent-ready · confidence high**


Assessment: A large sparse cousin with 6B active parameters for agentic coding and multimodal work, previewing Qwen's next architecture. It needs a 128 GB Mac for Ollama's MLX build or about 144 GB for the 4-bit GGUF, and Qwen Cloud serves it. The main risk is a licence that needs a separate deal for hosted-model and coding or office assistant businesses.

## Facts

- Kind: Open-weight model · vendor: Alibaba (Qwen team) · category: Open-weight models · legal entity: Alibaba Cloud (Singapore) Private Limited, the registrant of qwen.ai. The licence names Qwen as copyright holder · provenance 49/100
- Packages: huggingface `Qwen/Qwen3.8-Flash-Next`, huggingface `Qwen/Qwen3.8-Flash-Next-FP8`, ollama `qwen3.8-flash-next`
- Auth: None · pricing: Free · x402: no · licence: Qwen Community License 1.0 (custom, not an OSI licence). Free of charge, the model name shown in products above 100M monthly users or US$20M monthly revenue, and a separate licence from Qwen before commercial use by a Model as a Service or AI Work Assistant business
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/Qwen/Qwen3.8-Flash-Next and Qwen/Qwen3.8-Flash-Next-FP8, read from the model card, the files and their metadata, the chat template, the licence, Qwen's GitHub repository and technical report, and the runtimes' own code and recipes on 21 October 2026. Nothing was downloaded or run. The hosted Qwen3.8-Flash on Qwen Cloud, which the card says is based on this model with 1M context by default and built-in tools, isn't graded
- Model: Mixture of experts with 125B parameters and 6B active per token (512 experts, 10 routed and 1 shared), plus a 51B n-gram embedding table and a 4B multi-token prediction head, 180B in the files. 48 layers, three Gated DeltaNet layers to one Qwen Sparse Attention layer, and a vision encoder. Text, images and video in, text out. The card calls it an experimental preview of the architecture Qwen4 will use
- Files: 131 safetensors shards in BF16 (360.0 GB stored) with a chat_template.jinja, and a separate FP8 repository (185.5 GB). No pickle files and no Python files. Hugging Face lists a SHA-256 for every shard
- Licence: Qwen Community License 1.0, copyright Qwen. An MIT-style grant with two conditions. Products above 100M monthly active users or US$20M monthly revenue must show the model name prominently, and a licensee whose group runs a Model as a Service or AI Work Assistant business needs a separate licence from Qwen before commercial use, internal use excepted. No acceptable use policy in the repository
- Runtimes: Serving commands for SGLang, vLLM and TokenSpeed in Qwen's GitHub README, with recipes from vLLM and SGLang linked from the card. Model code (qwen4_exp) in Transformers 5.20.0, vLLM 0.30.0, SGLang 0.5.22 and mlx-vlm, the qwen4exp architecture in llama.cpp from release b11771, and a local qwen3.8-flash-next entry in Ollama's library. mlx-lm has no model for it
- Context: 262,144 tokens native. The card says it extends to 1,000,000 with YaRN, set in the config or on the command line, and warns that static YaRN can cost quality on short texts
- Tools and reasoning: The chat template declares tools in `<tools>` and expects calls as `<function=...>` and `<parameter=...>` blocks inside `<tool_call>`, which the `qwen3_coder` parser reads, as Qwen's README serving commands set. The card doesn't describe tool calling. Thinking is on by default, `enable_thinking: false` turns it off, `reasoning_effort` is xhigh (default), medium or low, and earlier thinking is kept unless `preserve_thinking` is false
- Hardware: Alibaba states no memory needs. The repositories hold 360.0 GB in BF16 and 185.5 GB in FP8, per Hugging Face. vLLM's recipe gives 172.78 GiB for FP8 and 335.28 GiB for BF16, and Ollama's smallest build is 105 GB
- Training data: Not described on the card. The technical report of 26 August 2026 covers architecture, ablations and training stability, and says only that the model trained on about a third of the tokens of its 397B-A17B predecessor, with no sources, mix or cut-off

- Scores: Capability 74, Reach 65, Risk 45, Agent interface 70, Documentation and upkeep 36 · total over the 7 assessed categories
- Why: Capability, From 7 independent benchmark results in 4 areas, each scored as its share of the way from the benchmark's floor to the best result any model… · Reach, Memory 5 of 30, the smallest listed 4-bit build is Ollama's 120 GB q4_K_M, 144 GB with a fifth added, which is beyond the reference machine… · Risk, File safety 25, safetensors only and nothing from the repository runs to load it or build prompts. · Agent interface, Chat template 20, a Jinja file in the repository. · Documentation and upkeep, Model card 11 of 20, intended use, architecture, parameters and context are there, but limitations, languages and a knowledge cut-off aren't…
- Sources: 21, open questions: 5, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/qwen3-8-flash-next.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/qwen3-8-flash-next.svg` or a link to https://www.anchorterminal.com/tools/qwen3-8-flash-next from a page on qwen.ai 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. On a 128 GB Mac start it with `ollama run qwen3.8-flash-next:125b-mlx` (105 GB, NVFP4), on Linux the smallest Ollama build is `qwen3.8-flash-next:125b-a6b-q4_K_M` at 120 GB, which won't fit a 107.4 GB GPU share
2. On a server, `vllm serve Qwen/Qwen3.8-Flash-Next-FP8 --tensor-parallel-size 4 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder`
3. Use the qwen3_coder tool parser and the qwen3 reasoning parser, as Qwen's GitHub README does
4. Thinking is on by default, pass `chat_template_kwargs` with `enable_thinking` false to skip it, and read the reasoning from `reasoning_content`, even though SGLang's cookbook says thinking can't be turned off
5. `reasoning_effort` takes xhigh (default), medium or low, and any other value makes the template raise an error
6. Send at most one system message and put it first, since the template raises an error otherwise
7. Sampling per the card is temperature 1.0 and top_p 0.95 when thinking, temperature 0.7, top_p 0.8 and presence_penalty 1.5 without
8. vLLM keeps the 51B n-gram table in host memory on 80 GB GPUs by default, per its recipe
9. Validate JSON output yourself or use the runtime's guided decoding, since structured output isn't documented
10. Read the Qwen Community License before serving it to others as an API or inside a coding or office assistant product

## Connect

```bash
ollama run qwen3.8-flash-next:125b-mlx
# or, as Qwen's GitHub README serves it (4 GPUs):
vllm serve Qwen/Qwen3.8-Flash-Next --port 8000 --tensor-parallel-size 4 --max-model-len 262144 \
  --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder
```

```bash
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen3.8-Flash-Next",
    "messages": [{"role": "user", "content": "Hello!"}],
    "chat_template_kwargs": {"enable_thinking": false}
  }'
```

## 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, desk reviews from public material, no calls made)

- ★★☆☆☆ An FP8 config fix in August, then one reply in 64 (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Custom licence, 185.5 GB at FP8 (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ An experimental preview with conflicting guidance (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Safe load path, a licence that limits who may use it (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, and all four give 2. They agree the files are clean but that Alibaba documents and answers little for an experimental preview, and two flag a licence that needs a separate deal for model-as-a-service and AI work assistant businesses. Read it as a large preview of Qwen's next architecture, best reached through Qwen Cloud by anyone whose business falls outside the licence's limits.