Alibaba Hasn't Shipped Qwen4 Yet — Its Preview Model Already Triggered a Quantization Race

Alibaba's Qwen team released Qwen3.8-Flash-Next on Hugging Face and ModelScope on August 26 under the Qwen Community License, explicitly framing it as an architectural preview rather than a finished product: a 125-billion-parameter mixture-of-experts model with only 6 billion active parameters per token, a 51-billion-parameter n-gram embedding table that can offload to system RAM, Gated DeltaNet attention across three of every four layers, and a native 262,144-token context window extensible to 1 million via YaRN. One early reviewer at eesel.ai called it "an intentionally under-trained R&D preview," distinct from the hosted Qwen3.8-Flash API Alibaba says is still coming, and further still from the full Qwen4 family the architecture is previewing.

That caveat hasn't slowed the ecosystem that has grown up around it. According to machine.net's Hugging Face activity tracker, comparing October 1 against September 30, the single fastest-growing model repository on the platform by download percentage was ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF, up 345% in one day — from 33,259 downloads to 148,142. A second quantization of the same base model, ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF, grew 51% in the same window, from 63,537 to 96,094. Both are built on GSQ-RCO, a quantization format ISTA-DASLab (the Distributed Algorithms and Systems Lab associated with the Institute of Science and Technology Austria) published for the model shortly after its release.

A format, not just a model, is what's spreading

What distinguishes this surge from an ordinary popular-model trend is that GSQ-RCO itself is propagating across independent publishers, not just one lab's download count climbing. Hugging Face listings show at least three separate accounts — ukisai, taurusduan, and pfeifferj — have each published their own GSQ-RCO re-quantizations of Qwen3.8-Flash-Next within weeks of ISTA-DASLab's original release, each offering slightly different variant sets (IQ2_XS, IQ3_S, IQ3_XXS, Q2_0, and multimodal-projector builds among them) aimed at running the 125-billion-parameter model on consumer GPUs with 24GB of memory or less.

"The quantized weights are already published, so this looks like a natural next step for the long-context local inference this project enables." — xiaoliboyone, in a GitHub feature request asking kvmem-llama.cpp maintainers to add native Qwen3.8-Flash-Next support

That request, filed as issue #96 against the kvmem fork of llama.cpp, asks for a support roadmap and timeline and offers to contribute testing on RTX 3090 (sm_86) hardware — the kind of unpaid tooling labor that typically follows a model only after it has proven durable demand, not one Alibaba itself still describes as a preview. The gap between what Qwen3.8-Flash-Next is officially positioned as and how the open-source community is already treating it — as production-worthy infrastructure worth building inference support around — is itself the story: it suggests that for local-inference practitioners, a well-documented quantization format from a credible lab can now outpace the official release cadence of the model it's built on.