The short version: Qwen3.8-Flash-Next is the first public look at Alibaba's Qwen4 architecture, and it is one of the most hardware-interesting models of 2026. On paper it is a ~180GB, 125-billion-parameter model, but only 6 billion parameters are active per token, and a chunk of it (a 51GB lookup table) can be streamed from an SSD. That combination means people are running it on a 48GB Mac, and even CPU-only, at around 12 tokens a second. The catches: the quants that fit a normal machine are 1 to 2-bit, and the license is not Apache. Here is what it is and what runs it.
We have not tested this model. This draws on the model card, a verified technical report, the license text, and owner reports, all linked. Alibaba's benchmark figures are the maker's claims.
What Qwen says it is: 125B active-6B, plus a 51B lookup table
Released August 26, 2026, this is billed as a preview of the Qwen4 architecture. The parameter accounting is unusual: a 125B backbone with only 6B active per token, plus a separate 51B "n-gram embedding" table and a 4B multi-token-prediction head, which is why the files on disk total around 180GB. It is multimodal, with a 262,144-token native context extensible to a million. The claimed benchmarks are frontier-adjacent: SWE-bench Pro 62.5, LiveCodeBench v6 91.9, AndroidWorld 84.5, each edging out the dense Qwen3.8-27B in Alibaba's own table. Alibaba also claims it "substantially reduces both training and inference cost, training takes only about 1/9 as much" as Qwen3.7-Plus. Read the per-benchmark numbers as claims, but here they are backed by an actual paper.
The Qwen4 architecture in plain English
Three ideas make this model cheap to run for its size, and all three are documented in the verified technical report (arXiv:2608.30320, "On the Design of Qwen3.8-Next Architecture," 36 authors, submitted August 31). First, the 48 layers use mostly Gated DeltaNet linear attention with a sparse-attention layer every fourth block, so long context stays cheap. Second, it is a sparse Mixture-of-Experts with 512 experts but only 10 routed plus 1 shared active per token, hence the tiny 6B active footprint. Third, and this is the clever hardware trick, the 51B n-gram table. In Alibaba's words, "N-gram Embedding looks up a table using the local context to scale model capacity with very little extra computation; the embedding table can be offloaded to host memory." A big slice of the model can live in slow memory or on disk because it is only ever looked up, never streamed through the GPU. That is what makes the 48GB-Mac demo possible.
What actually runs it
Because only 6B is active and the n-gram table can be offloaded, this model shows up on machines that have no business running a 180B checkpoint. Real, attributed reports:
- 48GB Mac (M5 Pro): a Show HN demo (slotstream, a Swift/MLX binary) keeps the small dense trunk resident and streams the 4-bit experts from SSD, hitting "~12 tok/s" warm decode with about 32GB peak memory.
- CPU-only (Core Ultra 9, 95GB RAM, no GPU): about 11 tok/s on the 1 to 2-bit builds.
- DGX Spark (NVFP4): about 25 tok/s decode, 77 tok/s prompt.
- Strix Halo (Ryzen AI Max+ 395): around 20 tok/s at low quant.
- 4x RTX 5090: 84 tok/s at 4.5K context, but decaying to 32.5 tok/s by 93K, a reminder that speed falls with depth until the speculative-decoding path is everywhere.
The quant ladder (bartowski GGUF): IQ1_S 70GB, IQ4_XS 97.7GB, Q4_K_M 119.6GB, Q8_0 188GB; the MLX 4-bit build is 111.6GB. Practically, a 96GB machine runs 1 to 2-bit, a 128GB machine runs the IQ4_XS, and a good 4-bit needs 128GB-plus with the SSD-offload trick. Confirm the fit in our Can I run it? calculator and pick the file with the quant picker.
The license: not Apache
Worth reading before you build on it. Unlike the dense Qwen3.8-27B (Apache 2.0), this ships under the Qwen Community License 1.0. It grants broad rights to use, modify, host, fine-tune, and even sell, which covers almost every local and internal use. But it is not Apache, and it carries a vaguely worded clause around "Model as a Service" and "AI Work Assistant" products that a commercial deployer should read. We map the 2026 open licenses, including this one, in our model license guide. For a home user the practical answer is: fine.
What the community says
Two themes dominate. The first is a plea for a smaller version: the top-voted discussion is simply "35B-A3B or 35B-A5B Humanity needs this MoE model," because the current builds are too big for a 16 to 32GB machine. As one commenter summed up the frustration, "Even at q4 you need more than 64 GB RAM. Most systems have 16-32 GB RAM." The second is that quality holds up better than expected at 4-bit but wobbles at 1 to 2-bit, which is exactly where budget machines are forced to run it; one Strix Halo owner found the IQ4_XS build "decent (very similar performance to qwen3.8 27b)." So in practice this is a model for 128GB-plus enthusiasts today, with the smaller sibling still on the wishlist.
Our take
Qwen3.8-Flash-Next is more exciting as a preview of where local models are going than as a daily driver. The architecture, tiny active set plus an offloadable memory table, is exactly what makes giant models tolerable on unified-memory boxes, and the fact that it runs at all on a 48GB Mac is the headline. If you have a 128GB machine and enjoy the frontier, grab the IQ4_XS and try it, but wait for the runtimes to finish landing the speculative-decoding path, and wait for the 35B version if your machine is smaller. This is the shape of Qwen4, seen early.
Sources and how we researched this
- Model card, architecture, claims: Qwen/Qwen3.8-Flash-Next on Hugging Face; technical report arXiv:2608.30320; license text Qwen Community License 1.0.
- Quant sizes from the bartowski and Unsloth GGUF repositories; the 48GB-Mac and CPU-only runs, DGX Spark and Strix Halo numbers, and the community threads from the Hugging Face discussions and Hacker News. We have not tested this model first-hand.
Related: The attention rebuild · Why active parameters decide what runs · The 2026 model license map · M5 Max vs M5 Ultra for local LLMs
