Qwen 3.8-Flash-Next releasing tomorrow (125B a6B)
Alibaba’s Qwen3.8-Flash-Next model is drawing attention as a large MoE (mixture-of-experts) architecture with 125B parameters and 6B active per token, designed to run on high-end “consumer” hardware like 128GB Macs, Strix Halo systems, and GPUs with 32GB+ VRAM. Commenters see it as a potential Opus/Sonnet-class local alternative for coding and “agentic” workloads, but note trade-offs in speed, memory footprint, quantization quality, and the practicality of running such models at home versus using cloud APIs. Many view this release as an early technical preview of the upcoming Qwen4 family, intended to let inference stacks and tooling catch up before the full lineup arrives.
Model architecture & goals
- Early description (since removed) said: 125B-parameter multimodal MoE, ~6B parameters active per token, plus ~51B n‑gram embeddings and a new “Qwen Sparse Attention.”
- Claimed to reach near‑Qwen3.7‑Plus capability at ~1/9th training cost, with better coding/cowork performance.
- Qwen team says this is mainly an architectural preview so inference stacks can prepare for the upcoming “full family” of Qwen4 models, not a fully polished release.
Parameter counts, scaling, and “effective size”
- One rule-of-thumb discussed: an MoE’s “effective dense size” ≈ geometric mean of total and active params. For 125B‑a6B this ≈27B, matching expectations that its quality will be similar to Qwen3.8 27B.
- Some expect capabilities roughly in the Sonnet/Opus‑4.6 class for coding and Linux tasks, but with extra “overthinking” that can be mitigated by temperature and prompting.
Hardware requirements & performance
- Many assume 128GB RAM class machines (Mac Studio, Strix Halo, DGX Spark, big GPUs) are the real targets.
- Several benchmarks and anecdotes:
- Qwen3.8 27B runs well on 5090 and dual‑GPU desktops; 32GB+ VRAM and ≥6‑bit quant preferred to avoid quality loss.
- Strix Halo struggles with 27B dense models (often ~10–30 tok/s), but MoE with 6B active may land around 25–40 tok/s; prefill latency remains a pain.
- M5 Max shows big gains over M4 for prefill; reported 25–70 tok/s on 27B with MTP and optimized MLX runtimes, though others see closer to 30–35 tok/s.
- FP8 for the full model seems too large for 96GB VRAM; advice is to expect FP4/Q4 for consumer setups.
Local vs cloud, and routers
- Many enjoy high‑end local models but find them slow for interactive coding/agent workflows vs. paid cloud APIs.
- OpenRouter praised for flexibility but criticized for:
- Qwen endpoints being flaky or capacity‑limited.
- Worse prompt‑caching economics vs. going direct to providers.
- Difficulty pinning to stable, cheap providers without lots of configuration.
- Some prefer self‑hosted routers (e.g., with model aliases, local/cloud fallbacks) for predictability and cost control.
Comparisons and ecosystem impact
- Model is expected to compete with DeepSeek‑style “flash” MoEs and other 27–35B‑class models (Qwen3.6/3.8, Gemma 4, Laguna, Ornith, GPT‑OSS).
- Mixed sentiment: excitement about a powerful, semi‑consumer‑runnable MoE; skepticism over speed, RAM prices, and whether it truly replaces Claude/other frontier APIs.