# Qwen3-Coder

Qwen3-Coder is an AI model released by Qwen on Jul 22 2025. It has 480B parameters, a 256k context window and open weight. At release it scored 20% on BullshitBench v2.

## Facts

| Field | Value |
| --- | --- |
| Model | Qwen3-Coder |
| Developer | Qwen |
| Release date | Tuesday, Jul 22 2025 |
| Licensing | Open Weight |
| Parameters | 480B |
| Context window | 256k |

## API pricing

All rates in USD per 1,000,000 tokens, pay-as-you-go.

| Tier | Input | Output |
| --- | --- | --- |
| Up to 32K input tokens | $1.00 | $5.00 |
| Over 32K and up to 128K | $1.80 | $9.00 |
| Over 128K and up to 256K | $3.00 | $15.00 |
| Over 256K and up to 1M | $6.00 | $60.00 |

Deployment: International (Singapore).
Verified August 18, 2026 against the first-party source: https://www.alibabacloud.com/help/en/model-studio/model-pricing

## Benchmark scores published at release

| Benchmark | Score | Source | What it measures |
| --- | --- | --- | --- |
| BullshitBench v2 | 20% | [BullshitBench](https://github.com/petergpt/bullshit-benchmark) | Given a confidently-worded but nonsensical prompt, does the AI spot that it makes no sense and push back — instead of playing along and inventing an answer? The score is how often it clearly called out the nonsense. Higher is better. |

## About Qwen3-Coder

Qwen3-Coder, released July 22, 2025, was Alibaba's largest open-weight model to that point: a 480B-parameter mixture-of-experts model with 35B active, built specifically for agentic coding rather than code completion. It shipped with a 256K-token native context window, extensible to roughly 1M through positional extrapolation, so a whole repository could sit inside a single prompt.

It arrived alongside Qwen Code, a command-line agent forked from Google's Gemini CLI, which signalled what the model was for — long multi-step tool-using sessions rather than single-shot completions. Released the same month as Kimi K2 and GLM-4.5, it formed part of a mid-2025 cluster of Chinese open-weight agentic models that reset what freely downloadable weights were expected to do, and a smaller Qwen3-Coder-Flash followed for latency-sensitive work.

## Questions and answers

### When was Qwen3-Coder released?

Qwen3-Coder was released by Qwen on Tuesday, Jul 22 2025.

### Who made Qwen3-Coder?

Qwen3-Coder was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud.

### How much does Qwen3-Coder cost?

Qwen3-Coder costs $1.00 per million input tokens and $5.00 per million output tokens through the Qwen API. Those are the rates for the “Up to 32K input tokens” tier; 3 other pricing tiers are published for this model. Rates are pay-as-you-go API prices verified against Qwen's published pricing on August 18, 2026.

### What benchmark scores did Qwen3-Coder get?

Qwen3-Coder reports 1 tracked benchmark score — BullshitBench v2: 20%. Scores are the figures published at release by Qwen.

### What is the context window of Qwen3-Coder?

Qwen3-Coder has a context window of 256k. That is the maximum amount of input plus output the model can hold in a single request.

### How many parameters does Qwen3-Coder have?

Qwen3-Coder is reported at 480B parameters.

### Is Qwen3-Coder open source?

Partly. Qwen3-Coder is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense.

### What came before and after Qwen3-Coder?

Qwen's previous tracked release was Qwen3 on Apr 29 2025, 84 days earlier. It was followed by Qwen3-Next on Sep 11 2025.


---

Canonical page: https://aireleasetracker.com/model/qwen/qwen3-coder
Full dataset: https://aireleasetracker.com/llms-full.txt · JSON: https://aireleasetracker.com/models.json
Source: AI Release Tracker (https://aireleasetracker.com). Most benchmark scores come from lab launch material; gathered results identify the leaderboard that published them.