Qwen 4 Is Announced; “Qwen 4 Coder” Is Not
Until 22 September 2026, Alibaba had never announced a "Qwen 4" generation. After Qwen 3.5 and 3.6, the version number jumped straight to 3.8 — and much of the community, assuming "4 must be next", filled the gap with rumor posts: a "Qwen 4.1", a "Qwen 4 Preview", and the subject of this page, a "Qwen 4 Coder 32B-A3B" complete with a June release date, an Apache 2.0 license, and an 82% SWE-bench Verified score. None of it was ever published through the Qwen team's blog, GitHub, or Hugging Face organization — the channels that carry every genuine Qwen release within hours.
The Apsara announcement changed only the first part of that story: Qwen 4 now exists as a model in training. None of the circulated specs came from Alibaba, and no Coder variant has been named.
During a July 2026 data incident, LLMCheck ingested several of those unverified entries, this one included. An August 2026 reconciliation against official sources confirmed that no “Qwen 4” of any kind had been published, and the entries were removed from the leaderboard and benchmarks. The full story of the phantom names is in the companion correction, Qwen 4 vs Qwen 4.1: What Exists and What to Run.
How the Fake Specs Gave Themselves Away
In hindsight, the circulated spec sheet had three tells worth learning from:
- No weights, anywhere. Every real Qwen release ships with a model card and downloadable weights on the official Hugging Face organization, usually mirrored to Ollama and LM Studio within days. "Qwen 4 Coder" existed only in aggregator tables and forum posts. If you cannot pull it from an official repo, it is not a release.
- The benchmark was too good. The claimed 82% SWE-bench Verified — from a 32B model running on a laptop — would top every vendor-reported open-weight score: 80.2% for Inkling-Small (Thinking Machines, bash-only harness) and 81.6% on Beam 1.0’s card (harness not stated). Inkling-Small has 276B parameters and needs a 96 GB+ Mac. Frontier-beating numbers from mid-size models are the single most reliable rumor signature.
- The specs were borrowed. A ~32B-class MoE with ~3B active parameters under Apache 2.0 is essentially the real Qwen3.6-35B-A3B with the version number filed off and the scores inflated.
The good news: you lose nothing by the phantom not existing. The verified coding models below are real, downloadable today, all Apache 2.0 — and the best of them is within a few points of the invented spec sheet anyway.
The Verified Mac-Runnable Coders
Here is the verified state of open coding models that fit a normal Mac, as of October 2026:
| Model | Architecture | Coding score (vendor-reported, vendor’s harness) | License | Mac to run it |
|---|---|---|---|---|
| Qwen3.6-27B | Dense 27B | 77.2% SWE-bench Verified (Alibaba, internal scaffold) | Apache 2.0 | 24 GB (16–19 GB at 4-bit) |
| Qwen3.6-35B-A3B | MoE 35B, 3B active | 73.4% SWE-bench Verified (Alibaba, internal scaffold) | Apache 2.0 | 24 GB tight, 32 GB comfortable (20–24 GB at 4-bit) |
| Qwen3-Coder-Next | MoE 80B, 3B active | 70.6% SWE-bench Verified (Alibaba, SWE-Agent) | Apache 2.0 | 64 GB (45–52 GB at 4-bit) |
| KAT-Coder-V2.5 | MoE 35B, 3B active | 69.4% SWE-bench Verified (Kwaipilot, Claude Code harness) | Apache 2.0 | 24 GB tight, 32 GB comfortable (19–22 GB at 4-bit) |
| Qwen3.8-27B | Dense 27.8B | 61.7% SWE-bench Pro (Alibaba, Claude Code harness) | Apache 2.0 | 24 GB (18 GB at 4-bit) |
| “Qwen 4 Coder” | — | Claimed 82% (never published by Alibaba) | — | Not released |
Every score above is reported by the model’s own developer on its own harness, so the column ranks models only roughly. SWE-bench Pro is a harder test than SWE-bench Verified, and the two are not comparable; Alibaba publishes no Verified score for Qwen3.8-27B. Sizes are the published 4-bit files (MLX, GGUF or Ollama).
The quality ceiling is not a coding specialist at all: the dense generalist Qwen3.6-27B posts 77.2% on SWE-bench Verified. Speed runs the other way. According to the LLMCheck index, on an M5 Max the mixture-of-experts coders are three to four times quicker: Qwen 3.6-35B-A3B at sourced118 tok/s (a published llama.cpp Q4_K_M run), KAT-Coder-V2.5 at community89 tok/s (a community MLX 4-bit run) and Qwen3-Coder-Next at estimated~84 tok/s, against estimated~30 tok/s for Qwen3.6-27B and estimated~29 tok/s for Qwen3.8-27B. The gap widens on smaller Macs: on a Mac mini M6, Qwen 3.6-35B-A3B runs at community64 tok/s and Qwen3.8-27B at community9 tok/s (both community-measured). The mark beside each figure shows whether it is published, community-measured or estimated. Per-chip figures for Qwen3.6-27B: M5 Max, M4 Max, M5 Pro; or see the broader roundup in Best Local LLM for Coding on Mac.
But dense 27B tokens are expensive, and agentic coding burns tokens by the thousand. That is where the two dedicated coders come in.
Qwen3-Coder-Next: the Real Coding Qwen (80B-A3B)
Qwen3-Coder-Next is the coding specialist Alibaba actually shipped, in early 2026: a mixture-of-experts model with 80B total parameters and 3B active per token, released under Apache 2.0. Alibaba reports 70.6% on SWE-bench Verified with SWE-Agent (model card), and because per-token compute tracks the small active set, generation runs in the same speed class as the far smaller A3B Qwens rather than a dense 80B.
The catch is footprint: the full expert pool has to sit in memory even though only slivers of it fire per token. The 4-bit builds are 45 GB in MLX and 52 GB in Ollama’s Q4_K_M, which rules out 32 and 48 GB machines in practice and makes a 64 GB Mac the home — with room left for a long context, your editor, and the OS. On a 64 GB Mac mini M5 Pro the index estimates estimated~49 tok/s. Per-chip figures: M5 Max, and every other Mac in the Best-by-Mac rankings. For setup details and quant choices, see the dedicated guide: Qwen3-Coder-Next on Mac.
KAT-Coder-V2.5: the 24–32GB Pick (35B-A3B)
KAT-Coder-V2.5, from Kwaipilot, arrived in July 2026 and is the most interesting budget entry in the category. Its open-weight “Dev” release is a 35B-A3B mixture-of-experts model post-trained from Qwen3.6-35B-A3B specifically for agentic coding, under Apache 2.0 (model card). Kwaipilot reports 69.4% on SWE-bench Verified with a Claude Code harness, and its 4-bit builds are 19–22 GB.
That footprint is the point: the 18.8 GB IQ4_XS GGUF fits a 24 GB Mac with a lean context, and 32 GB runs it comfortably, with the fast per-token generation of a 3B active set. A community run measured community89 tok/s on an M5 Max and community74 tok/s on an M4 Max; the index estimates estimated~66 tok/s on an M4 Pro. If your workload is tool-calling agent loops rather than one-shot questions, this is the strongest dedicated coder a 24–32 GB Mac can hold. Per-chip figures: M6, M4 Max, M5 Max.
Run KAT-Coder-V2.5 if…
You have 24–32 GB of unified memory and run coding agents. 69.4% SWE-bench Verified (Kwaipilot’s harness) in a 19–22 GB 4-bit build is the best capability per gigabyte among dedicated coders, and the 3B-active MoE keeps long agent loops fast.
Run Qwen3-Coder-Next if…
You have 64 GB and want specialist headroom — 70.6% SWE-bench Verified (Alibaba, SWE-Agent) with the same fast A3B generation, plus a much larger expert pool for harder multi-file work. On 48 GB-and-under machines it does not fit with room to work; do not force it.
What to Run by RAM Tier
- 8–16 GB — Skip this class. The honest option is Bonsai 27B, Prism ML’s 1-bit derivative of Qwen3.6-27B (Apache 2.0, about 3.9 GB; 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. Or see the best model for a 16 GB Mac.
- 24 GB — KAT-Coder-V2.5 (IQ4_XS) or Qwen 3.6-35B-A3B (MLX 4-bit, 20 GB) with a lean context. Both are tight; close memory-heavy apps.
- 32–48 GB — Qwen 3.6-35B-A3B or KAT-Coder-V2.5 for fast agent loops; Qwen3.8-27B or Qwen3.6-27B when answer quality matters more than speed.
- 64 GB — Qwen3-Coder-Next for coding, with a 3B-active generalist alongside. This tier runs both.
- 96–128 GB — Inkling-Small (80.2% SWE-bench Verified, Thinking Machines’ bash-only harness), estimated at estimated~42 tok/s on a 128 GB M5 Max at 2-bit (Q2_K).
Every model recommended here is Apache 2.0, and so is Qwen3.8-27B, which shipped on 14 August 2026 and is the strongest dense model in the LLMCheck index by capability (46/50; Score 71, estimated~29 tok/s on an M5 Max). Not every new Qwen is: Qwen3.8-Flash-Next ships under the Qwen Community 1.0 license, which requires a separate commercial license for coding-assistant products. Qwen 4 has no published licence yet; this page will say which one it gets when weights appear.
Install & Setup
Every model here is on Hugging Face, and the Qwens are in Ollama’s library under the tags below (checked 18 October 2026). Ollama 0.40 and later runs model architectures that MLX supports on MLX by default on Apple Silicon (release notes):
# Speed pick for 24–32 GB Macs (about 24 GB download)
ollama run qwen3.6:35b-a3b
# Dense quality pick (18 GB download)
ollama run qwen3.8:27b
# Coding specialist for 64 GB Macs (52 GB download)
ollama run qwen3-coder-next
# KAT-Coder-V2.5 is not in Ollama's library; pull the GGUF from Hugging Face
ollama run hf.co/bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF:IQ4_XSIn LM Studio, search the model name and pick an MLX 4-bit or Q4_K_M build that fits your memory. The community run behind KAT-Coder-V2.5’s M5 Max figure used the 22 GB mlx-community OptiQ 4-bit build. The per-chip model pages linked above carry the current LLMCheck figure for each Mac.