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_effortat 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 the dense Qwen3.8-27B, 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 mlx-community 4-bit conversion is 112 GB, the ggml-org IQ4_NL GGUF is 102 GB and its Q8_0 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 practical rule: Qwen3.8-Flash-Next is a 128 GB-tier model. A 96 GB Mac does not hold a 4-bit build. A 128 GB Mac holds the 112 GB MLX build with about 16 GB for context, which is workable, and the 102 GB GGUF with a little more. On a 192 GB or 256 GB Studio the constraint disappears.
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 well below models with lower capability scores: the 128 GB floor costs accessibility points by design.
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. Each figure below is the index’s current one for that Mac, refilled on every build: measured where a plain-decode run is linked, otherwise estimated from the nearest measured chip. The mark on each figure says which.
| Mac | Bandwidth | Index tok/s |
|---|---|---|
| Mac Studio / MacBook Pro M5 Max, 128 GB | 614 GB/s | community44 tok/s |
| Mac Studio / MacBook Pro M4 Max, 128 GB | 546 GB/s | estimated~40 tok/s |
| Mac Studio M2 Ultra, 192 GB | 800 GB/s | estimated~54 tok/s |
| Mac Studio M3 Ultra, 256/512 GB | 819 GB/s | estimated~55 tok/s |
| Mac Studio M5 Ultra, 256/512 GB | 1,228 GB/s | community57 tok/s |
Put next to its neighbours on a 128 GB M5 Max — DeepSeek V4 Flash at community39 tok/s, GLM-5.3-Flash at estimated~27 tok/s, Qwen3.8-27B at estimated~29 tok/s — Flash-Next is the fastest capable model at the tier, which is why the index’s per-Mac pages pick it first on the M3, M4 and M5 Max 128 GB configurations.
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.
The coding-assistant clause is the one to notice. Building a local coding tool for yourself is exactly what the license permits. Shipping one to customers on this model is exactly what it gates. The LLMCheck index scores the license as Qwen Custom, 5/10, against Apache 2.0’s 8/10 — and if you need Apache, Qwen3.8-27B is the same generation at 19 GB.
How to run it
The card names SGLang, vLLM and TokenSpeed for servers. On a Mac there are three routes. Ollama 0.33.1 and later runs it on MLX: ollama run qwen3.8-flash-next:125b-mlx (105 GB, so it needs a 128 GB Mac; the 120 GB q4_K_M tag leaves too little headroom). The other two are below.
llama.cpp with GGUF
Support for the qwen4exp architecture has been in llama.cpp since 27 August 2026, so use a current build. The ggml-org repo carries IQ4_NL (102 GB), Q2_K (83 GB) and Q8_0 (163 GB), plus a 617 MB vision projector and separate multi-token-prediction heads:
llama-server -hf ggml-org/Qwen3.8-Flash-Next-GGUF:IQ4_NLThis is the route behind the 57 tok/s M5 Ultra figure. Because the model is multimodal, load the mmproj file alongside the weights if you want image input.
MLX
pip install -U mlx-vlm
mlx_vlm.generate --model mlx-community/Qwen3.8-Flash-Next-4bit \
--prompt "Summarise this repository" --max-tokens 512A 112 GB download. mlx-vlm 0.7 or later loads it; stock mlx-lm does not yet (pull request #1788 is still open). The oMLX server and MTPLX also run it; MTPLX produced the 44 and 47 tok/s M5 Max figures above.
Whichever route, set the context deliberately. With about 16 GB free on a 128 GB machine, 32K is comfortable and the 1M ceiling is theoretical.
Frequently Asked Questions
How much RAM does Qwen3.8-Flash-Next need on a Mac?
128 GB. Its 4-bit builds are 102 GB (the IQ4_NL GGUF) to 112 GB (MLX), because the model's 51B n-gram embedding table is held in unified memory alongside the 125B of expert weights; the Q8_0 GGUF is 163 GB. A 96 GB Mac cannot hold a 4-bit build. The LLMCheck index lists a 120 GB minimum, the 128 GB tier.
How fast is Qwen3.8-Flash-Next on Apple Silicon?
As of 18 October 2026 the LLMCheck index holds community-measured plain-decode runs (multi-token prediction off): 44 tok/s on an M5 Max (the middle of three runs: 47.0 and 43.8 in MTPLX, 28.5 in oMLX), about 27 on an M3 Ultra (26.5 and 25.1 measured) and 57 on an M5 Ultra (llama.cpp IQ4_NL). Other chips are estimated from the nearest measured one, about 40 tok/s on an M4 Max. Only 6B parameters stream per token, which is why it outpaces every other capable model at the 128 GB tier.
Can I use Qwen3.8-Flash-Next commercially?
For internal and local use, yes, without restriction. Under the Qwen Community 1.0 license, offering it as a model-as-a-service or building an AI Work Assistant product for coding or office productivity requires a separate commercial license from Alibaba, and products above 100 million monthly active users or US$20 million monthly revenue must display the model name prominently. It is not Apache 2.0.
Is Qwen3.8-Flash-Next better than Qwen3.8-27B?
On the model card's benchmarks, slightly: SWE-Bench Pro 62.5 vs 61.7, LiveCodeBench v6 91.9 vs 90.3, GPQA Diamond 91.7 vs 89.2. On a 128 GB M5 Max it is also faster: 44 tok/s community-measured against about 29 estimated for the 27B. Qwen3.8-27B wins on everything else: it fits a 24 GB Mac at 18 GB, and it is Apache 2.0. The LLMCheck index ranks the 27B well above Flash-Next (#12 against #38), because RAM tier and license count.
Is Qwen3.8-Flash-Next Qwen 4?
No. Its model card calls it an experimental preview of the architecture that will underpin Qwen4 (the Gated DeltaNet plus Qwen Sparse Attention hybrid and the n-gram embedding), released as a 3.8-generation model. Alibaba announced on 22 September 2026 that Qwen 4 itself is in training; as of 18 October 2026 there are no Qwen 4 weights.