Best Qwen model for coding: cloud and local, by job

The best Qwen model for coding in the cloud is qwen3.8-max, which Alibaba lists for complex reasoning and coding. For cheap, fast edits use qwen3.8-flash; for a balance, qwen3.7-plus. Locally, qwen3.8:27b is the pick with 24 GB of memory, qwen3.5:9b with 12 GB, and qwen3-coder-next with 64 GB or more. These names were checked on 5 November 2026.

Which cloud Qwen model for which job?

Alibaba’s own Claude Code setup for Token Plan uses this split: Max for the Opus slot, Flash for Sonnet, and qwen3.6-flash for Haiku. That is a good pattern for any agent with a strong-model and a cheap-model slot.

The same Model Studio page also lists models you don’t want for code, such as the omni, audio, image and embedding models. A name that contains omni or audio is not a coding model.

The retired-name trap: Alibaba retires models. Ones marked “Upcoming deprecation” in the Model Gallery return a 403 AccessDenied. If an old config suddenly fails, check the name against the current list before you suspect your key. On Coding Plan, only the exact versions Alibaba lists are supported.

What about long context?

Qwen3.8-27B has 262,144 tokens of native context, and its card says it can reach 1,000,000 with YaRN scaling, with a warning that this may hurt performance on shorter texts. For the hosted models, Alibaba’s Token Plan Claude Code setup sets CLAUDE_CODE_MAX_CONTEXT_TOKENS to 983616, and says that for 1M you set it to 1000000 or add [1m] to supported model names. Alibaba’s page doesn’t say which models those are, so test one before you rely on it.

A longer context costs more per request, because the agent resends it every turn. See AI coding costs.

Which Qwen model for coding runs locally?

Open weights are Apache-2.0 for Qwen3.8-27B and Qwen3-Coder-Next. These tags, sizes and contexts were checked against Ollama’s library, and they all carry the tools label:

The download isn’t the memory you need: the context lives beside the model. The sizes at 64k and 128k, the Ollama context default and the tags for 8 GB are all in best Ollama models for coding.

Other tags worth knowing:

• qwen3-coder:30b (19 GB): the older Coder model, no thinking, fast on small turns.
• qwen3-coder:480b (290 GB at 4 bits): for a server, not a laptop.
• qwen3.5:27b (17 GB) and qwen3.5:35b-a3b (22 GB): the previous generation.
• qwen3.8:27b-mlx (18 GB): built for Apple silicon.

How good are they?

Qwen reports these numbers on the model cards, run by Qwen in its own setups:

The two rows use different benchmarks and different versions, so don’t subtract one from the other. AI coding benchmarks, explained says why. Only your own repo settles it: give each model one small fix and one change across several files.

Which agent runs them?

Qwen Code, Claude Code, Codex CLI, OpenCode, Cline and Kilo Code all connect: see the Qwen page for the cloud setup and the Ollama page for local. Which plan covers which model is in Qwen Coding Plan and Token Plan.

Contenox

Contenox lists the cloud models above in its Model chip once the Alibaba Cloud Model Studio provider is connected, and the local ones through Ollama, so one session can switch between qwen3.8-max in the cloud and qwen3.8:27b on your GPU. Setup: Qwen coding agent.

Related guides

• Best Ollama models for coding: all local picks by VRAM
• Qwen Coding Plan and Token Plan: which models each plan includes
• Alibaba Cloud Model Studio for coding agents: endpoints and keys
• How to choose a model for coding: test on your own tasks