Qwen3.8-Flash-Next: a 6B-active preview of Qwen4, and the weights are public

On August 26 Alibaba released Qwen3.8-Flash-Next, an open-weight multimodal MoE it explicitly bills as "an early preview of the architecture used in Qwen4": a 125B-parameter main model plus 51B of n-gram embeddings, with just 6B parameters activated per token. The hosted production version, Qwen3.8-Flash, shipped the same day on Alibaba Cloud at $0.15 in / $0.47 out per 1M — which finally closes the "Qwen publishes no readable price" gap we have carried since June.

Verdict: Open weights, 125B total but only 6B active — the Qwen4 architecture preview lands at the top of the consumer range

The take

The facts, from the model card (`Qwen/Qwen3.8-Flash-Next`) and the Model Studio page: 125B language model + 51B n-gram embedding + 4B MTP layer — 180B parameters in the safetensors, BF16, so ~360 GB unquantized. Context is 262K native, extensible to 1M; the hosted Qwen3.8-Flash serves 1M by default. Vision input is native. Alibaba says training cost about 1/9th of Qwen3.7-Plus while beating it on coding and office tasks — vendor numbers, treat accordingly. Within two days the card showed 4,810 downloads and 92 quantized variants, so the ecosystem moved on it immediately.

⚠️ The license is `qwen-community-1.0` — a custom community license, not Apache 2.0. That matters here specifically: Qwen3.8-Max already shipped revenue-gated in exactly this pattern, and "open weights" in headlines tells you nothing about what you may build on commercially. Read the LICENSE file on the card before treating this as a Qwen3.8-27B-style permissive drop. We have not yet done a clause-level read; until we have, we label it "open weights, restricted license — unconfirmed scope."

Why the shape is interesting for local hardware: 6B active parameters means the compute per token is small — the bill is memory, not FLOPs. A Q4 quant of 180B lands somewhere near 100 GB (our arithmetic, not a measured figure), which is out of reach of every consumer GPU but inside a 128 GB Mac Studio or a multi-channel EPYC/Strix-Halo box with room to spare. If the quality holds up under quantization, this is the most consumer-plausible frontier-adjacent MoE since the DeepSeek V4-Flash class — and much lighter per token than that 284B/13B-active model.

Our call: no planner-pick change yet, for three reasons in order. The license scope is unconfirmed; the quant quality is unmeasured (92 quants existing is adoption, not evidence they are good); and no mainstream local runner had a blessed recipe on release day. What would change our mind: a confirmed benign license read, a llama.cpp/Ollama/MLX path with real tok/s reports on 128 GB Macs, and the Q4 holding within a few points of BF16 on coding evals. The hosted route needs no such caveats: $0.15 / $0.47 per 1M with a published USD rate card is now a row in our cost calculator — the first Qwen row ever, after three sweeps of finding their pricing page 404.

Where this fits

Models: Qwen3.8-27B · Qwen3.8-Max (2.4T-A95B) · DeepSeek V4-Flash

Hardware: Mac Studio M3 Ultra 96 GB · Mac Studio M4 Max 64 GB

Sources

Next step

Try this in the planner→