Qwen 4 vs Qwen 4.1: What Exists, and Which Qwen to Run on a Mac

What Happened to "Qwen 4"?

Qwen 4 was announced on 22 Sep 2026 and is still in training, with no open weights; Qwen 4.1 does not exist. Before that announcement, nothing named Qwen 4 had been announced at all. After the Qwen 3.5 and 3.6 releases, much of the community assumed the next major version would be called "Qwen 4", and rumor posts filled in the blanks: a "Qwen 4.1 32B-A3B", a "Qwen 4 Coder", even a "Qwen 4 Preview". The invented specs were plausible precisely because they mirrored a real model — the circulated "32B-class MoE, ~3B active parameters, Apache 2.0, 256K context" description is essentially Qwen3.6-35B-A3B with the version number filed off.

During a July 2026 data incident, LLMCheck ingested several of those unverified entries into its catalog. An August 2026 reconciliation against official Qwen channels (the team's blog, GitHub, and Hugging Face organization) confirmed that no "Qwen 4" of any kind had been published, and the entries were removed from the leaderboard and benchmarks.

The Real Roadmap: 3.5 → 3.6 → 3.8

Here is the verified state of the Qwen line as of 24 October 2026. Coding scores are vendor-reported on the vendor’s own harness; SWE-bench Pro is harder than SWE-bench Verified and the two are not comparable.

Two structural notes. First, the jump from 3.6 to 3.8 is Alibaba's own numbering: Qwen 3.7 exists only as hosted models (Alibaba's model cards use Qwen3.7-Max and Qwen3.7-Plus as baselines), so there was no 3.7 open release, which is part of why outsiders kept guessing "4 is next". Second, the 3.8 flagship comes in two forms: open weights (Qwen3.8-2.4T-A95B) and a hosted version with extra features (Qwen3.8-Max).

Qwen3.6: What to Run on a Mac

For local use, the Qwen3.6 family is the story, and it gives you a clean two-way choice.

Qwen3.6-27B — the capability pick

Qwen3.6-27B is a dense 27B model and one of the strongest open coder-generalists that fits on a normal Mac: Alibaba reports 77.2% on SWE-bench Verified. It was the LLMCheck Mac #1 in August 2026 (Score 72) and scores 70 today, at estimated~30 tok/s on an M5 Max. Being dense, every token pays the full 27B cost, so it wants memory bandwidth: the index estimates estimated~15 tok/s on an M5 Pro and estimated~9 tok/s on an M6. A 24GB Mac runs the 16–19 GB 4-bit build with a modest context. Per-chip figures: M5 Max, M4 Max, M5 Pro, M2 Max.

Qwen3.6-35B-A3B — the speed pick

Qwen3.6-35B-A3B is the mixture-of-experts sibling: 35B total parameters with 3B active per token, under Apache 2.0. It gives up a few benchmark points (Alibaba reports 73.4% on SWE-bench Verified) but generates far faster than the dense 27B, because per-token compute tracks the 3B active set: sourced118 tok/s on an M5 Max (a published llama.cpp run) and community64 tok/s on a 32 GB Mac mini M6 (community-measured). If your Mac has 24–32GB, or your workload is long agentic loops where token throughput compounds, this is the one. Per-chip figures: M5 Max, M4 Max, M6.

Qwen3.8: Max, the 2.4T Open Weights, the 27B and Flash-Next

The 3.8 generation arrived in August 2026 in four parts, two of which run on a Mac.

• Qwen3.8-Max is Alibaba’s hosted version of the 2.4T model, with extras such as a 1M-token default context. It is API-only — no weights — so it sits outside the scope of the LLMCheck index.
• Qwen3.8-2.4T-A95B open weights arrived in August: 2.4 trillion total parameters with 95B active. Alibaba’s model card reports 67.7 on SWE-bench Pro (Claude Code harness) and 92.6 on GPQA Diamond for the hosted version. Two catches. The smallest published GGUF is a 1-bit file of about 397GB, and 4-bit is about 1.3TB, so it is server-class. And it ships under a custom license with revenue thresholds — the first Qwen 3.x open release not under Apache 2.0.
• Qwen3.8-27B shipped on 14 August 2026 under Apache 2.0: a 27.8B dense vision-language model that runs at 4-bit in about 19GB, so it fits a 24GB Mac. It is the strongest dense model in the LLMCheck index (Score 71), at estimated~29 tok/s on an M5 Max. Every figure for it.
• Qwen3.8-Flash-Next, which its model card calls an experimental preview of the architecture that will underpin Qwen4: 125B parameters with 6B active, a 128GB-tier model under the Qwen Community 1.0 license, at community46 tok/s on an M5 Max. Flash-Next on a Mac.

Recommendations by Mac Tier

• 8–16GB — Skip the 27B class. Qwen3.5-9B is the sensible Qwen at this tier, at estimated~18 tok/s on a base M4 (per-chip figures); Prism ML’s 1-bit Bonsai 27B (3.9 GB, a Qwen3.6-27B derivative) is the alternative. Or see the Best by Mac pages for your exact machine.
• 24GB — Qwen3.6-35B-A3B at 4-bit with a short context (tight), or the dense Qwen3.8-27B (an 18 GB download).
• 32GB — Qwen3.6-35B-A3B comfortably, the LLMCheck index’s #1 and its best pick at 32GB or less; Qwen3.8-27B when you want the strongest dense model, with image and video input.
• 64GB — the same picks with room for long context; for coding, the 80B-A3B Qwen3-Coder-Next becomes an option at this tier.
• 128GB — Qwen3.8-Flash-Next, the Qwen 4 architecture preview (check its license before building products on it).

Installing the Real Qwens

The Mac-class Qwens are in Ollama’s library under these tags (checked 24 October 2026). Ollama 0.40 and later runs supported models on MLX by default on Apple Silicon:

In LM Studio, search “Qwen3.6” or “Qwen3.8” and choose the quant that fits your memory tier. The mlx-community 4-bit conversions are 16 GB for the 27B models and 20 GB for the 35B-A3B. Step-by-step setup is in the guide to running Qwen on a Mac.