Qwen3.8-Flash-Next (125B-A6B preview)
Experimental preview of the Qwen4 architecture: 125B MoE with 6B active (plus 51B n-gram embedding and 4B MTP). 262K native context. Thinks by default. Big-memory box territory.
Recommended quants
- Unsloth UD-Q4_K_XL (111.3 GB files)
unsloth/Qwen3.8-Flash-Next-GGUFFits 128 GB unified-memory machines only with small context. - Unsloth UD-Q2_K_XL (78.9 GB) / UD-IQ1_S (72.5 GB)
unsloth/Qwen3.8-Flash-Next-GGUFLow-bit options for 96 GB-class memory; quality loss unmeasured here. - Q8_0 (199.3 GB) / BF16 (367 GB)
unsloth/Qwen3.8-Flash-Next-GGUFMulti-GPU or very large Macs.
Chat template Works
Qwen chat template. Thinking on by default (<think>...</think>). Controls: enable_thinking, preserve_thinking (both on by default), reasoning_effort.
Fix: For direct answers pass chat_template_kwargs {"enable_thinking": false}. The production version with 1M context by default is the API-only Qwen3.8-Flash.
Sampling settings
| Thinking (official) | temp 1.0, top_p 0.95, top_k 20, min_p 0, presence_penalty 0, repetition_penalty 1.0 |
|---|---|
| Instruct / non-thinking (official) | temp 0.7, top_p 0.80, top_k 20, min_p 0, presence_penalty 1.5 |
Runtimes
| vLLM | Unverified | |
|---|---|---|
| llama.cpp | Works | Unsloth GGUFs published (new hybrid QSA attention: use a recent build). |
| Ollama | Unverified | Library page qwen3.8-flash-next exists; local vs cloud tag not checked. |
| LM Studio | Unverified | |
| MLX | Unverified |
Tool calling in agent tools
| Claude Code (local endpoint) | Unverified | |
|---|---|---|
| OpenCode | Unverified | |
| Cline | Unverified | |
| Aider | Unverified |
Known issues
Sources checked
- https://huggingface.co/Qwen/Qwen3.8-Flash-Next
- https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF
- https://ollama.com/library/qwen3.8-flash-next
Spotted something wrong or a new fix? Corrections are reviewed before publishing. (Submission form coming soon.)