What Alibaba Has Confirmed, and What It Hasn’t
Until November, everything Alibaba had said about Qwen 4 fitted in one paragraph of its 22 September 2026 press release from the Apsara Conference in Hangzhou. One hosted preview has followed:
- Confirmed: Qwen 4 is “currently in training”. The same release puts Qwen 4.5 and Qwen 5 on a roadmap “projected to scale up to 5 to 10 trillion parameters”.
- Since then: Qwen4-Max-Preview, a hosted model on Alibaba Cloud Model Studio from 11 to 25 November 2026 (UTC+8), through the Token Plan only and retired when the preview ends; 1M-token context, text, image and video input (Model Studio documentation). It confirms that a Max tier exists, and nothing else: no weights, licence or size.
- Not confirmed: a release date, a licence, model sizes, any tier other than the hosted Max preview, and whether any Qwen 4 model will ship as open weights. Third-party accounts of the keynote name tiers and sizes; none of them appears in the press release.
- Not on Hugging Face: the Qwen organization has no Qwen 4 repository (checked 14 November 2026). Its newest language models are Qwen3.8-27B and Qwen3.8-Flash-Next, both from August. The “Qwen4-27B” repositories other accounts have created hold no weights.
How this tracker decides. Every genuine open-weight Qwen release has appeared on the Qwen Hugging Face organization with a model card and weights. Qwen 4 joins the LLMCheck leaderboard when that happens, and not before; a size or tier joins the table above when Alibaba publishes it.
Qwen3.8-Flash-Next: the Qwen 4 Architecture, Today
You can already run the architecture Qwen 4 will use. Qwen3.8-Flash-Next, released on 26 August 2026, is described on its model card as an “experimental preview of the architecture that will underpin Qwen4”: 125B parameters with 6B active per token, plus a 51B n-gram embedding and a 4B multi-token-prediction head, with Gated DeltaNet paired with Qwen Sparse Attention, native image and video input and a 262K-token context.
On a Mac it is a 128 GB-tier model, because the n-gram table sits in memory with the experts and the 4-bit MLX build is about 112 GB. According to the LLMCheck index it runs at community47 tok/s on a 128 GB M5 Max, the lower median of eight community-measured plain-decode runs, all with multi-token prediction off: 43.8 and 47.0 tok/s in MTPLX 2.10, 46.4 in llama.cpp, 50.1 in Ollama, 52.6 in mlx-vlm 0.7.10, 58.7 in oMLX 0.7.2 and 59.6 in oMLX 0.7.3.dev2, plus an older oMLX 0.6.3 run at 28.5, and community57 tok/s on an M5 Ultra (a community llama.cpp IQ4_NL run). With multi-token prediction on, the same MTPLX builds reach 73.5 and 75.9 tok/s on the M5 Max; that is a different workload, so the index keeps it out of the speed rows.
Runtime support is in place on the llama.cpp side: the qwen4exp architecture was merged on 27 August (PR #27742) and multi-token-prediction drafting on 1 October (PR #29761). On MLX it runs through mlx-vlm 0.7 or later and oMLX; the stock mlx-lm pull request (#1788) was still open on 14 November 2026. Its Qwen Community 1.0 licence is the catch for builders; see Qwen3.8-Flash-Next on a Mac for setup and the clause to read first.
What Qwen 3.8 Actually Is
Qwen 3.8 shipped in August 2026, not the way the “Qwen 4” rumour mill predicted. Its flagship is Qwen3.8-2.4T-A95B, a 2.4-trillion-parameter mixture-of-experts model with 95B active parameters, released as open weights; Qwen3.8-Max is Alibaba’s hosted version of it, with extra features such as a 1M-token default context. A Mac-class Qwen3.8-27B followed on 14 August under Apache 2.0, and Qwen3.8-Flash-Next on 26 August. There was no open Qwen 3.7: Alibaba’s own model cards use Qwen3.7-Max and Qwen3.7-Plus, both hosted, as baselines.
| Model | Status | Role | Licence |
|---|---|---|---|
| Qwen3.8-Max | Hosted | API and Qwen Chat, built on the 2.4T weights | Hosted only — no download |
| Qwen3.8-2.4T-A95B | Weights out, August 2026 | Open-weight flagship, server-class | Custom, revenue thresholds |
| Qwen3.8-27B | Shipped 14 Aug 2026 | Mac-class dense model (24 GB Macs) | Apache 2.0 |
| Qwen3.8-Flash-Next | Shipped 26 Aug 2026 | Qwen 4 architecture preview (128 GB Macs) | Qwen Community 1.0 |
| Qwen3.6-27B / 35B-A3B | Shipped April 2026 | Current Mac picks | Apache 2.0 |
The open-weight flagship posts serious numbers: on its model card, Alibaba reports 67.7 on SWE-bench Pro (Claude Code harness) and 92.6 on GPQA Diamond for Qwen3.8-Max. But the smallest published GGUF is a 1-bit file of about 397 GB, and the 4-bit build is about 1.3 TB. This is a server-class release.
According to the LLMCheck index, two Qwen 3.8 models run on Macs: Qwen3.8-27B (24 GB) and Qwen3.8-Flash-Next (128 GB). The index’s best pick at 32 GB or less is still a Qwen 3.6 model, Qwen 3.6-35B-A3B (Score 81).
The Licences
Qwen 3.x open releases had all shipped under Apache 2.0 until August. Qwen3.8-2.4T-A95B is the first to use a custom licence instead. Its terms add two obligations: products above 100 million monthly active users or US$20 million in monthly revenue must display the model name prominently, and a model-as-a-service or AI work-assistant business whose revenue tops US$50 million over any 12 months needs a separate licence from Qwen. Qwen3.8-Flash-Next goes further: its Qwen Community 1.0 licence requires a separate licence for any model-as-a-service or coding and office assistant product, whatever its revenue.
- For individuals and most teams, neither licence bites. Running a model locally, or using it inside your company without exposing it to third parties, is permitted.
- For companies shipping products, read the terms before building on Qwen 3.8 weights. If you want no conditions at all, Qwen3.8-27B and the Qwen 3.6 line remain Apache 2.0.
The leaderboard scores licence openness as part of every model’s total: Apache 2.0 earns 8 of 10 points and the custom Qwen licences 5.
Qwen3.8-27B: Shipped 14 Aug 2026
The release Mac users cared about most, Qwen3.8-27B, shipped on 14 August 2026: a 27.8B dense vision-language model with image and video input, under Apache 2.0. Ollama’s build is an 18 GB download, so it fits a 24 GB Mac, and it is the strongest dense model in the LLMCheck index by capability (46/50; Score 72). Press reviews measured it at 46–55 tok/s on an M5 Ultra; the index lists sourced48 tok/s there, community31 tok/s on an M5 Max and community10 tok/s on a Mac mini M6. Dense 27B models are slow on base chips, so on a 24–32 GB Mac most people will prefer the faster 35B-A3B. Every figure for it →
What Mac Users Should Run Today
What to run now, by unified memory, according to the LLMCheck index:
- 8–16 GB — Bonsai 27B or Qwen 3.5 9B. Prism ML’s Bonsai 27B is a 1-bit derivative of Qwen3.6-27B: about 3.9 GB, Apache 2.0, with vendor-reported ~90% of full-precision quality and sourced44 tok/s on an M5 Pro; it needs Prism ML’s llama.cpp or MLX fork. Qwen 3.5 9B runs on stock tools at estimated~18 tok/s on a base M4. The index’s overall pick for 16 GB is a Gemma model; see best by RAM.
- 24–32 GB — Qwen 3.6-35B-A3B. It activates 3B of its 35B parameters per token, is tied for the index’s #1 at Score 81 (MiMo-V2.6-Flash, a 192 GB-class model, is listed first on the alphabetical tie-break) and is its best pick at 32 GB or less: community130 tok/s on an M5 Max and community59 tok/s on a 32 GB Mac mini M6. On 24 GB it is tight; keep the context short. Qwen3.8-27B is the slower, stronger dense alternative.
- 64 GB — add Qwen3-Coder-Next for coding. The 80B-A3B coding specialist runs at estimated~49 tok/s on a 64 GB M5 Pro. The verified Qwen coders.
- 128 GB — Qwen3.8-Flash-Next. The Qwen 4 architecture preview, community47 tok/s on an M5 Max. Check its licence before building products on it.
ollama run qwen3.6:35b-a3b # 3B-active MoE, 24 GB+ Macs (about 24 GB download)
ollama run qwen3.8:27b # dense 27.8B, 24 GB+ Macs (18 GB download)
ollama run qwen3-coder-next # coding specialist, 64 GB Macs (52 GB download)
# Tags checked on ollama.com on 14 November 2026
New to local models? Start with the guide to running Qwen on a Mac or the Ollama install guide. The best LLM by Mac pages rank every model that fits your exact chip and RAM.
Run a current Qwen if…
You want the best Qwen on a Mac today. On 24–32 GB, Qwen 3.6-35B-A3B is the index’s best pick, tied for #1 overall at Score 81 (MiMo-V2.6-Flash is listed first on the alphabetical tie-break); Qwen3.8-27B is the strongest dense model by capability (46/50), Apache 2.0, with image and video input. Neither depends on Qwen 4’s timing.
Wait on Qwen 4 if…
You need Qwen 4 specifically. Alibaba has given no date, sizes or licence, and nothing announced so far makes the current Qwen picks obsolete on Apple Silicon. This tracker updates when weights appear.