Qwen 3.8-Flash-Next
A 125B MoE architecture preview of Qwen4 with only 6B active parameters per token
Context window
262K
Input / 1M tokens
Free
Output / 1M tokens
Free
Provider
Alibaba
Data verified 2026-08-27
Qwen3.8-Flash-Next is an open-weight multimodal mixture-of-experts model that serves as an early preview of the Qwen4 architecture. It features 125 billion total parameters with only 6 billion activated per token, plus 51 billion parameters in an N-gram embedding layer. The model introduces architectural innovations including hybrid attention (Gated DeltaNet + Qwen Sparse Attention), ultra-sparse MoE, and multi-token prediction, targeting cost-efficient long-context and agentic workloads.
Capability index
Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.
How to access it
Available as open-weight on Hugging Face and ModelScope. A production API version (Qwen3.8-Flash) is served on QwenCloud at $0.16/$0.47 per million input/output tokens, but the open-weight preview itself has no published API pricing.
Strengths
- ✓Extremely cost-efficient with 6B active parameters out of 125B total
- ✓Open-weight for local deployment
- ✓Native multimodal (text and image input)
- ✓256K native context window, extensible to 1M tokens
- ✓Optimized for agentic and long-context tasks
Best for developers who...
When to choose it (and when not to)
Reach for Qwen 3.8-Flash-Next when...
- →When cost per token is a primary constraint
- →For on-premise or local deployment
- →When you need long context window support
- →For agentic or multi-step reasoning tasks
Look elsewhere if...
- ✕If you require official API support (open-weight version only)
- ✕If data residency outside China is a strict requirement
- ✕If you need the fully-released Qwen4 family (this is a preview)
How to use it
- ›Allocate sufficient output tokens for complex reasoning (up to 262K recommended for internal reasoning)
- ›Use preserved thinking mode for agent scenarios to maintain reasoning trace
- ›For texts exceeding 262K natively, consider RoPE scaling techniques like YaRN
Quickstart
Pythonfrom transformers import AutoTokenizer, AutoModelForCausalLM
model_id = 'Qwen/Qwen3.8-Flash-Next'
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map='auto')FP8 quantized version available for reduced memory requirements
API model id: qwen3.8-flash-next
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| DeepSWE 1.1 (code generation) | Not published | Evaluated with Claude Code and mini-SWE-agent harnesses at 256K context |
| Text Arena | Comparable to Qwen3.8-Max flagship | Outperforms DeepSeek-V4-Flash and Claude Opus 4.6 on majority of tasks |
Source: Qwen Official Blog
Compare Qwen 3.8-Flash-Next
Compare Qwen 3.8-Flash-Next with any other model
Build a comparison →All model comparisons →