Models

Qwen Coder: Qwen3-Coder, Qwen3-Coder-Next, Qwen 3.8 and the Qwen Code CLI

Which Qwen model to use for coding now, how each scores, what runs on your computer, what the API costs, and what changed with the Qwen Code CLI.

Published 2026-10-29 · by Richard Kaminsky and Mitchell Lipyansky

"Qwen Coder" is Alibaba's line of coding models: Qwen3-Coder-480B-A35B, released in July 2025, and Qwen3-Coder-Next, released February 3, 2026, with 80 billion parameters (3 billion active), a 262K-token context and the Apache 2.0 license. There's been no newer Coder model since, and Alibaba now recommends its general Qwen 3.8 models for coding instead: Qwen3.8-27B (August 14, 2026, Apache 2.0), an 18 GB download that runs on a 24 GB graphics card, and Qwen 3.8 Max, which scored 38.9% on Vals AI's independent Terminal-Bench 4.0 run against 65.2% for Claude Opus 5.5. Qwen Code, Alibaba's open-source terminal agent, no longer has a free tier: it ended on April 15, 2026.

Key takeaways

  • Coder models: Qwen3-Coder-480B-A35B (July 2025) and Qwen3-Coder-Next (Feb 2026, 80B, 3B active), both Apache 2.0.
  • Use now: Alibaba says to "use the latest general-purpose model instead of the Qwen-Coder model."
  • On your computer: qwen3.8:27b (18 GB) on 24 GB; qwen3-coder-next (52 GB) on a 64 GB Mac.
  • Best Qwen on independent tests: Qwen 3.8 Max, 38.9% on Terminal-Bench 4.0, 13th on LMArena WebDev.
  • Qwen Code CLI: Apache 2.0, about 28,800 GitHub stars; now runs on an API key, a coding plan or other providers.
Qwen's coding models, from Hugging Face, Ollama, Alibaba Cloud Model Studio, Vals AI (Oct 29) and LMArena (Oct 27).
ModelReleasedParameters (total / active)LicenseOllama sizeIndependent result
Qwen3-Coder-480B-A35BJul 2025480B / 35BApache 2.0290 GBLMArena WebDev #135
Qwen3-Coder-NextFeb 3, 202680B / 3BApache 2.052 GBNot ranked
Qwen3.8-27BAug 14, 202627B denseApache 2.018 GBSWE-bench Verified 86.0%; WebDev #34
Qwen 3.8 MaxAug 20262.4T / 95BQwen's ownCloud onlyTerminal-Bench 4.0 38.9%; WebDev #13
Claude Sonnet 5.5 (reference)Sep 28, 2026Not publishedClosed—Terminal-Bench 4.0 64.1%; WebDev #3

Sources: the Qwen3-Coder-Next and Qwen3.8 model pages, Alibaba Cloud Model Studio, Ollama, the Qwen Code repository, Vals AI and LMArena, opened October 29, 2026.

What are the Qwen Coder models?

  • Qwen3-Coder-480B-A35B (July 2025): 480 billion parameters, 35 billion active, a 262K context that stretches to 1M, Apache 2.0. Too large for most local machines: 290 GB in Ollama.
  • Qwen3-Coder-Next (February 3, 2026): "80B in total and 3B activated," 262,144-token context, Apache 2.0. Qwen's own figure is 70.6% on SWE-bench Verified. In Ollama it's qwen3-coder-next, 52 GB.
  • Qwen3-Coder 30B (qwen3-coder:30b, 19 GB): the smaller coder for a single graphics card.

Alibaba's Model Studio page for the coder models now says: "It is recommended to use the latest general-purpose model instead of the Qwen-Coder model" (Alibaba Cloud).

Which Qwen model is best for coding now?

The Qwen 3.8 family, released in August 2026:

  • Qwen 3.8 Max: 2.4 trillion parameters, 95 billion active, under Qwen's own license. 38.89% on Vals AI's Terminal-Bench 4.0 (16th of 58), level with GLM-5.3 as the best downloadable models after Kimi K3.1 Pro. 13th on LMArena's WebDev votes (1676, preliminary).
  • Qwen3.8-27B: dense, Apache 2.0, a 262K context extensible to 1M. 86.0% on Vals AI's archived SWE-bench Verified run, and 34th on LMArena WebDev. The one to run locally.
  • Qwen3.8 Flash Next: 125 billion parameters, 6 billion active, 24th on WebDev.

None comes close to Claude Opus 5.5 (65.2% on Terminal-Bench, first on WebDev) or Sonnet 5.5 (64.1%, third) on long agent tasks. Their case is price, openness and running privately (the best open-weight models for coding).

How do you run Qwen locally?

With Ollama, sizes at the default quantization (Ollama):

  • 24 GB graphics card or 32 GB Mac: ollama run qwen3.8:27b (18 GB) or qwen3-coder:30b (19 GB).
  • 64 GB Mac: qwen3-coder-next (52 GB; 85 GB at 8-bit).
  • More memory: qwen3.8:27b at full precision is 56 GB.

Leave memory for the context: a long coding session uses far more than the model file. Agents that work with a local Qwen include OpenCode, Cline, Qwen Code itself and Codex CLI with --oss (Claude Code alternatives).

How much does the Qwen API cost?

On Alibaba Cloud Model Studio's international (Singapore) region, per million input and output tokens (Alibaba Cloud):

  • qwen3.8-max: $2 and $6.
  • qwen3.8-27b: $0.50 and $3.
  • qwen3.8-flash: $0.15 and $0.47.
  • qwen3-coder-flash: $0.30 and $1.50 up to 32K tokens of input, rising to $1.60 and $9.60 for 256K to 1M.

The larger coder models, qwen3-coder-plus and qwen3-coder-next, are no longer on the international price list. Claude Sonnet 5.5 is $2 and $10 for comparison (Claude API pricing).

What is the Qwen Code CLI?

Alibaba's open-source terminal coding agent (GitHub): Apache 2.0, about 28,800 GitHub stars, "originally based on Google Gemini CLI v0.8.2" and developed separately since its first release.

  • Install: npm install -g @qwen-code/qwen-code@latest (Node.js 22 or later), or brew install qwen-code.
  • Sign-in: "The Qwen OAuth free tier was discontinued on 2026-04-15." It now runs on QwenCloud, an Alibaba Model Studio coding plan or API key, or another provider's key, including local models.
  • Features: subagents, plan mode and a sandbox, much like Claude Code.

Qwen vs Claude for coding

  • Hard agentic work: Claude Opus 5.5 or Sonnet 5.5, ahead by about 25 points on Terminal-Bench 4.0.
  • Private, unlimited, on your own computer: Qwen3.8-27B with Ollama.
  • Cheap API calls: qwen3.8-flash at $0.15 and $0.47, or Claude Haiku 5.5 at $0.10 and $0.50.
  • Fine-tuning: Qwen's Apache 2.0 models; Claude can't be downloaded.

Other open models: GLM 5, DeepSeek V4 vs Claude, Kimi K3 vs Claude. Free options of every kind: the best free AI for coding.

Claude, GPT, Grok and Muse in a shared room

Poly (usepoly.co) doesn't offer Qwen. It runs Claude, GPT, Grok and Muse in one browser room where a team works with the same agent, with a free model from each lab and nothing to install. What is Poly?

Common questions

What is Qwen Coder?

Alibaba's coding models: Qwen3-Coder-480B-A35B from July 2025 and Qwen3-Coder-Next from February 2026, an 80-billion-parameter model with 3 billion active, both under Apache 2.0. Alibaba now recommends its general Qwen 3.8 models for coding.

What is the best Qwen model for coding?

Qwen 3.8 Max on independent agent tests, at 38.9% on Vals AI's Terminal-Bench 4.0. To run on your own computer, Qwen3.8-27B, an 18 GB download in Ollama that fits a 24 GB graphics card.

Is Qwen Code free?

The CLI is free and open source, but its free Qwen OAuth tier ended on April 15, 2026. It now needs an API key, an Alibaba coding plan or another provider, or a local model through Ollama.

Is Qwen better than Claude for coding?

No. Claude Opus 5.5 scores 65.2% on Vals AI's Terminal-Bench 4.0 against 38.9% for Qwen 3.8 Max. Qwen's strengths are open weights, Apache 2.0 licenses and running locally.

How big is Qwen3-Coder-Next?

80 billion parameters with 3 billion active per token and a 262,144-token context. In Ollama it is a 52 GB download at the default quantization, so it needs about 64 GB of memory.

About the authors

Written by Richard Kaminsky and Mitchell Lipyansky, the co-founders of Poly. Poly is a multiplayer AI coding workspace: a shared room where your team works with one AI agent, together. It runs at usepoly.co: any member can approve the agent's changes, and every turn ends with a receipt naming who ran it and what it cost. Corrections: support@usepoly.co.