Qwen 4 Coder: Why You Can't Download It — and the Real Coding Qwens

Why There Is No "Qwen 4 Coder"

Alibaba has 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.

During a July 2026 data incident, LLMCheck ingested several of those unverified entries, this one included. An August 2026 reconciliation against official sources confirmed no "Qwen 4" of any kind exists, and the entries were removed from the leaderboard and benchmarks. The full story of the phantom "Qwen 4" line is in our companion correction, 'Qwen 4' vs Reality: The Actual Qwen Roadmap.

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 beat the actual open-weight record as of August 2026: 80.2%, held by Inkling-Small, a 276B-parameter model that needs a 128GB Mac Studio to run at all. 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 Verified Mac-Runnable Coders

Here is the verified state of open coding models that fit a normal Mac, as of August 2026:

The quality ceiling is not a coding specialist at all: the dense generalist Qwen3.6-27B posts 77.2% on SWE-bench Verified and, according to the LLMCheck index, holds a Score of 72 — the current Mac-runnable #1 — at an estimated ~40 tok/s on an M5 Max. If your machine has 32GB or more and you want the best answers, that is still the default; see the per-chip estimates for M5 Max, M4 Max, and M4 Pro, or 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: a mixture-of-experts model with roughly 80B total parameters and about 3B active per token, released under Apache 2.0. It scores 70.6% on SWE-bench Verified, 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. At 4-bit the weights come to roughly 43GB (est.), which rules out 32GB machines and makes a 64GB Mac the comfortable home — with room left for a long context, your editor, and the OS. Per-chip estimates: M5 Max, M4 Max, M4 Ultra. For setup details and quant choices, see our 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. It is a 35B-A3B mixture-of-experts model built on the Qwen3.6 base and post-trained specifically for agentic coding, released under Apache 2.0. At 69.4% on SWE-bench Verified it lands within striking distance of the dense Qwen3.6-27B — while quantizing to a Q4 footprint of roughly 20GB (est.).

That footprint is the point: it fits a 24GB Mac (tight, with a lean context) and runs comfortably on 32GB, with the fast per-token generation of a ~3B active set. If your machine is an M4 Pro-class 24–32GB and your workload is tool-calling agent loops rather than one-shot questions, this is the strongest dedicated coder you can hold. Per-chip estimates: M4 Pro, M4 Max, M5 Max.

What to Run by RAM Tier

• 8–16GB — Skip this class entirely. Bonsai 27B, the ternary derivative of Qwen3.6-27B (Apache 2.0, ~4–6GB checkpoint, vendor-reported ~90% quality retention), is the honest option at this tier, or see Best by Mac for your exact machine.
• 24GB — KAT-Coder-V2.5 at Q4 with a lean context, or Qwen3.6-35B-A3B if you also want a generalist.
• 32–48GB — Qwen3.6-27B for the best quality; KAT-Coder-V2.5 alongside it for fast agent loops.
• 64GB — Qwen3-Coder-Next for coding, Qwen3.6-27B for everything else. This tier runs both side by side.

One forward note: a Qwen3.8-27B has been teased for August 2026, with license and benchmarks unpublished. Until it appears on official channels with downloadable weights, treat any "Qwen3.8-27B benchmark" you see with the same skepticism this page exists to teach — the LLMCheck index will list it when it ships.

Install & Setup

All three verified models are on Hugging Face; exact registry tags vary, so search the model name in Ollama or LM Studio and pick a Q4_K_M (or MLX 4-bit) build:

# Quality pick — the current Mac #1 (32GB+)
ollama run qwen3.6:27b

# Budget coder for 24–32GB Macs
# search "KAT-Coder-V2.5" in LM Studio, or pull the
# Kwaipilot / mlx-community build from Hugging Face

# Coding specialist for 64GB Macs
# search "Qwen3-Coder-Next" in your registry

On M-series Macs, MLX-community conversions of all three typically squeeze extra tok/s out of the unified-memory path; the per-chip model pages linked above carry the current LLMCheck estimates for each quant.

Frequently Asked Questions