Qwen 3.8-Flash-Next releasing tomorrow (125B a6B)
- ID
- 17777
- Status
- summarized
- Published
- 25 Aug 2026, 7:49 PM
- Fetched
- 26 Aug 2026, 2:57 AM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://modelscope.cn/models/Qwen/Qwen3.8-Flash-Next
- Source URL
- https://hnrss.org/best
Summary
- Score
- 6.5
- Created
- 26 Aug 2026, 2:58 AM
- Tags
- Audience
- developersai_ml_learnersai_agent_usersvibe_coders
What happened
Qwen is releasing Qwen3.8-Flash-Next on 2026-08-26 at 15:00 UTC, a 125B-parameter MoE model with 6B active parameters, built on the next-generation Qwen4 architecture. Two variants will be available: the standard model and an FP8-quantized version, both on ModelScope. The team describes this as an early architectural preview ahead of the full Qwen4 model family.
Why it matters
If you build with open-weight LLMs, this is a concrete date to benchmark against: a 125B MoE with only 6B active params could offer strong quality-to-cost ratios for agent workloads. Plan to download and test both the standard and FP8 variants on release day to see if they beat your current Qwen3 or DeepSeek setup before committing.
Discussion angle
Compare the 125B-a6B MoE spec against current go-to open models like Qwen3 or DeepSeek V3 — is the active-param count low enough to run locally or does it still require cloud GPU, and does the FP8 variant change that calculus?