Qwen3.8-Flash-Next on a Mac: 6B Active, 112 GB, and the License Clause to Read First (2026)

What Qwen3.8-Flash-Next is

Alibaba released Qwen3.8-Flash-Next on 26 August 2026 and describes it, unusually plainly, as a preview of the architecture that will underpin Qwen4 rather than a finished flagship. From the model card:

• 125B total parameters, 6B activated per token, plus a 51B n-gram embedding and a 4B multi-token-prediction head.
• A hybrid attention stack — Gated DeltaNet alternating with Qwen Sparse Attention — across 48 layers, and 512 experts of which 10 routed and 1 shared fire per token.
• Native text, image and video input.
• 262,144-token context natively, 1M with YaRN.
• Thinking on by default, with reasoning_effort at xhigh, medium or low.
• Reported: SWE-Bench Pro 62.5 (vendor scaffold), LiveCodeBench v6 91.9, GPQA Diamond 91.7, OSWorld 2.0 19.4 binary / 52.3 partial.

Each of those benchmark figures is a little above Qwen3.8-27B, the dense model that currently tops the LLMCheck leaderboard, and the index scores Flash-Next’s capability at 47 to the 27B’s 46. The SWE-Bench Pro number is annotated as vendor-scaffold because vendor harnesses typically run 15–20 points above standardized ones.

Why 4-bit is 112 GB, not 60

A 125B model at 4-bit should be about 70 GB. The community MLX conversion Vontra/Qwen3.8-Flash-Next-MLX-oQ4 is 112 GB, and the official ggml-org Q8_0 GGUF is 163 GB. The difference is the 51B n-gram embedding table, which is part of the checkpoint and sits in unified memory with the experts. Only one row of it is read per token, so it costs almost nothing in bandwidth — but it costs everything in capacity.

The index lists it with a minimum of 120 GB, which places it in the same tier as DeepSeek V4 Flash and GLM-5.3-Flash, and is why it ranks #13 overall despite a capability score in the top ten: the 128 GB floor costs accessibility points by design.

Estimated speed per Mac

Six billion active parameters is where the model earns its name. Streaming roughly 3.4 GB per token, it is bandwidth-light, and the estimates below reflect that — scaled by the index’s MoE factor (0.42 of the raw active-bytes formula, fitted to the one community-measured MoE run in the index). They are estimates, labelled as such on every row.

Put next to its neighbours on a 128 GB machine — DeepSeek V4 Flash at 39 tok/s community-measured, GLM-5.3-Flash at ~27 estimated, Qwen3.8-27B at ~29 estimated — Flash-Next is the fastest capable model at the tier, which is why the index’s per-Mac pages now pick it first on every 128 GB-and-up configuration. The linear-attention half of its stack may make it run better than the factor predicts; the first measured run will tell.

The license clause to read first

Qwen3.8-27B is Apache 2.0. Qwen3.8-Flash-Next is not. It ships under Qwen Community 1.0, and the terms that matter are these:

• Local and internal use is unrestricted. Running it on your Mac, for yourself or inside your company, is fine. Redistribution and derivatives are allowed with attribution.
• Model-as-a-service needs a separate commercial license from Alibaba before you offer it.
• “AI Work Assistant” products for coding or office productivity need a separate commercial license. Internal tools that no third party accesses are exempt.
• At scale, prominent display. Products past 100,000,000 monthly active users or US$20,000,000 in monthly revenue must display the model name prominently.

How to run it

The card names SGLang, vLLM and TokenSpeed for production and says llama.cpp, Ollama and LM Studio work through community quantizations. On a Mac the two routes are:

MLX

pip install -U mlx-lm
mlx_lm.generate --model Vontra/Qwen3.8-Flash-Next-MLX-oQ4 \
  --prompt "Summarise this repository" --max-tokens 512

112 GB download. The native route on Apple Silicon and the one behind the estimates above.

llama.cpp / LM Studio with GGUF

The official ggml-org repo currently carries only Q8_0 at 163 GB plus a 617 MB vision projector — too large for a 128 GB Mac. Community 4-bit GGUFs exist; check the file size against your tier before pulling. Because the model is multimodal, load the mmproj file alongside the weights if you want image input.

Whichever route, set the context deliberately. With ~16 GB free on a 128 GB machine, 32K is comfortable and the 1M ceiling is theoretical.

Frequently Asked Questions