Qwen 3.6

Alibaba's Qwen 3.6 is two stories in one. The good one for open-source builders: a dense 27-billion-parameter model that beats the previous generation's 397B Mixture-of-Experts flagship on coding, and runs on a single 24GB GPU, under Apache 2.0. The other one, worth saying out loud: the actual flagship, Qwen3.6-Max, is closed and API-only, a quiet break from the fully-open Qwen3.5 era. Here is what you can download, what you cannot, and how to run the part that is open.

What is open, and what is not

Alibaba's Qwen team released two genuinely open-weight Qwen 3.6 models in April 2026, both Apache 2.0 and both already past five million downloads: Qwen3.6-27B, a dense model, and Qwen3.6-35B-A3B, a Mixture-of-Experts model with about 3 billion active parameters. Both take text, image, and video, call tools, and carry a 256K-token context that stretches to roughly a million with YaRN scaling. What did not ship as weights is Qwen3.6-Max, the roughly trillion-parameter flagship, which exists only behind Alibaba's API. So "Qwen 3.6 is open" is true for the models most people will actually run, and false for the top of the lineup. Do not let a headline conflate them.

The interesting model: a 27B that beats a 397B

The dense 27B is the one to watch. Its trick is architectural: it leans on Gated DeltaNet, a linear-attention design, for three of every four layers, with standard attention in the rest. (Linear attention scales better with context length, which is how a 27B model holds a near-million-token window without melting.) It adds multi-token prediction for faster generation and keeps a thinking mode on by default. The payoff, in Qwen's own numbers, is that this 27B edges out the previous-generation Qwen3.5-397B-A17B on several coding benchmarks while fitting on a single consumer GPU. Intelligence per parameter, not raw scale.

Benchmarks (Qwen-reported)

The numbers below are from Qwen's own launch materials for the 27B. They compare mainly against its predecessors and Claude 4.5 Opus; there is no independent reproduction, and cross-vendor comparisons to Kimi or DeepSeek come only from third-party aggregators, so treat those as community-reported.

Qwen-reported figures for the dense 27B, not independently reproduced. Comparisons beyond the Qwen3.5 family and Claude are third-party.

Limitations and gotchas

• The flagship is closed. Qwen3.6-Max is API-only; the open story stops at 27B and 35B-A3B.
• Benchmarks are Qwen-reported. Strong, but unverified, and the headline comparison is against Qwen's own older model.
• Vision adds friction for some runners: the vision projector file complicates plain Ollama GGUF use, so llama.cpp, vLLM, or SGLang are the safer paths.
• Thinking on by default means more output tokens; turn it off when you just want a quick answer.

Who should use it

If you want a capable, openly licensed model that runs on one 24GB GPU and is genuinely good at code and tool use, the Qwen3.6-27B is one of the best picks of 2026, and a clear upgrade path from Qwen3.5 for local builders. Reach for the 35B-A3B MoE if you have a bit more memory and want faster inference per token. Skip the closed Max unless you specifically need its ceiling and are fine with an API. For most self-hosters, the open 27B is the story.

Run it in about 10 minutes

The dense 27B fits in roughly 17GB at 4-bit, so a 24GB card or a recent Mac runs it.

# Quantized local run via the community GGUF. A 24GB GPU or 32GB Mac is comfortable.
ollama run hf.co/unsloth/Qwen3.6-27B-GGUF:Q4_K_M

# Or serve the full weights with vLLM and the long context
vllm serve Qwen/Qwen3.6-27B --max-model-len 262144 --reasoning-parser qwen3

Give it a real repository task with tools enabled; tool calling and the long context are where the 3.6 generation pulled ahead of 3.5. If you run a coding agent, point a harness like OpenCode at the local endpoint and let it work.

Sources and further reading

• Qwen3.6-27B model card (Hugging Face)
• Qwen3.6-35B-A3B model card (Hugging Face)
• Qwen3.6 on GitHub (Apache 2.0)
• MarkTechPost: the 27B benchmarks

Tested on: not independently tested. The Qwen3.6-27B is reported to run in about 17GB at 4-bit on a single 24GB GPU or a 32GB Mac; benchmarks are Qwen-reported and not independently reproduced, with cross-vendor comparisons drawn from third-party aggregators. The flagship Qwen3.6-Max is closed and was not evaluated.
Date checked: 2026-06-26