Qwen: Qwen3 Coder 30B A3B Instruct

Qwen

•

Qwen3 Coder 30B A3B Instruct

•

Released

July 2025

Qwen
Qwen3 Coder 30B A3B Instruct

Compact open-weight MoE coder with 3.3B active parameters, 262K-token context and agentic tool calling for repository-scale tasks.

Modality:

Text

model ID

qwen/qwen3-coder-30b-a3b-instruct

Start for free

View API Docs

Output Speed *

N/A

tok/s

Intelligence Index *

9.6

/ 100

Context Window *

160000

tokens

Input price

0.42

Anytoken

Output price

1.68

Anytoken

Agentic Coding With 3.3B Active Parameters and 262K Context Qwen3 Coder 30B A3B Instruct is Qwen's compact coding model in the Qwen3-Coder family, sitting below the 480B flagship. It is a mixture-of-experts model with 30.5B total and 3.3B active parameters, a native 262,144-token context window, and a function-call format built for agent tools such as Qwen Code and Cline. It runs in non-thinking mode only. It suits repository-aware code generation, editing and tool-calling agents where response speed matters more than deep reasoning. Call Qwen3 Coder 30B A3B through AnyAPI.ai

Performance

Small Active Footprint, Solid Coding Output, Limited Deep Reasoning

The model activates 8 of 128 experts per token, so only 3.3B of 30.5B parameters run per forward pass. Artificial Analysis measured an Intelligence Index of about 10, above the median of 6 for similar open-weight non-reasoning models, and 40.3% on LiveCodeBench. That suggests it handles routine code generation and tool-driven edits better than its size implies, without being a frontier reasoner. In production it fits high-frequency coding calls and agent loops; escalate harder multi-step problems to a larger model.

Benchmarks

Qwen3 Coder 30B A3B on Artificial Analysis: Coding, Agents and Speed

Artificial Analysis lists LiveCodeBench at 40.3% and Terminal-Bench Hard at 15.2%, which points to competent everyday coding but limited success on difficult terminal-agent tasks. IFBench is 32.7% and τ²-Bench Telecom 34.5%, so strict instruction following and multi-turn tool use are modest. Speed depends on the host: Amazon measured 118.9 tokens per second with 1.19s time to first token, Alibaba Cloud 97.6 tokens per second and 2.69s. Artificial Analysis now marks the model deprecated and points to Qwen3 Coder Next.

Output Speed

N/A

tok/s

Intelligence Index

9.6

/ 100

MMLU *

Broad world knowledge and problem-solving

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.

HLE *

Adherence to multi-step structured instructions.

LiveCodeBench *

Tool-calling reliability in long agentic loops.

Technical Specifications

What the model supports

The model accepts text and returns text only. Its context window is 262,144 tokens natively, and Qwen documents YaRN extension to 1M tokens, though hosted providers may not expose that. Context window is not output length: Amazon Bedrock lists a 16K maximum output, while other hosts advertise much larger limits, so confirm the cap for your route before generating whole files or large diffs. It runs in non-thinking mode only, so there is no reasoning budget to tune.

Verified Specifications —

Qwen3 Coder 30B A3B Instruct

Input modalities

Text

output modalities

Text

Context window

160000

tokens

Maximum output tokens

32768

Reasoning

Knowledge cutoff

July 2025

Pricing (standard)

0.42

AnyTokens in

1.68

AnyTokens out

Comparison

Qwen3 Coder 30B A3B or Qwen3-Coder 480B A35B for Coding Agents?

Both models come from the Qwen3-Coder family, are listed by Qwen as instruct models, and share a 256K native context length, so the decision is capacity versus footprint. The 480B model activates 35B parameters per token; the 30B model activates 3.3B. That gap should favor the larger model on harder tasks, though this page does not cite a head-to-head measurement. The practical question is whether your workload needs maximum agentic coding quality or can accept a lighter model.

Dimension

Qwen3 Coder 30B A3B Instruct

Qwen3 Coder 480B A35B Instruct

Context window *

160000

tokens

262144

tokens

Output speed *

N/A

tok/s

0.00

tok/s

Intelligence Index *

9.6

11.9

Input pricing

0.42

AnyToken

Output pricing

1.68

AnyToken

Knowledge cutoff *

July 2025

Choose Qwen3 Coder 30B A3B when requests are frequent and bounded: completions, single-file edits, routine tool calls, or self-hosting on a single GPU. Its smaller active footprint and lower cost matter at volume. Choose Qwen3-Coder-480B-A35B-Instruct when tasks involve harder multi-file reasoning or longer agent trajectories, or when you see failures on the smaller model and can accept greater compute per request. Test both on your own repository tasks before committing.

Limitations & Trade-offs

Where Qwen3 Coder 30B A3B Instruct falls short

Modest agentic benchmark results. Despite agentic positioning, Artificial Analysis reports 15.2% on Terminal-Bench Hard and 34.5% on τ²-Bench Telecom. Short, well-specified tool loops are realistic; long-horizon terminal or multi-turn workflows will fail more often. For autonomous agents running many dependent steps, a larger Qwen3-Coder model or a newer successor is the safer choice.

No thinking mode. The model is non-thinking only and cannot spend extra tokens on deliberate reasoning. Artificial Analysis reports 51.6% on GPQA Diamond, 29.0% on AIME 2025 and 3.8% on Humanity's Last Exam. Algorithm design, math-heavy code and architectural decisions are better routed to a reasoning model, while this one stays on direct code generation.

Long context is not long-range reasoning. The 262,144-token window is native, yet Artificial Analysis reports 32.7% on AA-LCR, below the 53% it lists for Qwen3 Next 80B A3B Instruct. Output caps also differ by host, with Amazon Bedrock listing 16K. Verify retrieval accuracy on your own large repositories, and confirm the output limit before relying on large generated files.

Aging model with a successor. Amazon Bedrock lists the model as Legacy, and Artificial Analysis marks it deprecated and recommends Qwen3 Coder Next. It is also text-only, with no image input. Artificial Analysis calls it comparatively expensive among similar open-weight non-reasoning models, and hosted rates vary by provider. For new long-lived projects, evaluate the successor alongside it.

Best-Fit Workloads

Where this model earns its place

‍

Pricing in anytokens via AnyAPI

Input

0.42

₳

Output

1.68

₳

Cache write

—

₳

Cache read

₳

Get API access

Read Docs

Integration

Access Qwen3 Coder 30B A3B Instruct via AnyAPI.ai

Access Qwen3 Coder 30B A3B Instruct through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate Qwen3 Coder 30B A3B Instruct without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

One API integration

Access Qwen3 Coder 30B A3B Instruct and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

Easy model switching

Test Qwen3 Coder 30B A3B Instruct against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

Flexible for production

Use Qwen3 Coder 30B A3B Instruct from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

Multi-model applications

Use Qwen3 Coder 30B A3B Instruct for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free

Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

What is the Qwen3 Coder 30B A3B Instruct context window?

The context window is 262,144 tokens natively, and Qwen says it can be extended to 1M tokens with YaRN. Context window is separate from maximum output, which varies by provider; Amazon Bedrock lists 16K output tokens. Hosted providers may not expose the 1M extension.

Does Qwen3 Coder 30B A3B Instruct have a reasoning mode?

No. The model supports non-thinking mode only and does not generate thinking blocks, so there are no reasoning controls to configure. It answers directly, which suits code generation and tool calls. For math-heavy or multi-step problem solving, use a dedicated reasoning model instead.

Does Qwen3 Coder 30B A3B Instruct support tool calling and images?

It supports tool calling through a specially designed function-call format, and OpenRouter lists tools and tool_choice parameters. It does not support images: input and output are text only. Structured JSON schema output is listed by OpenRouter, but availability depends on the hosting provider.

How does Qwen3 Coder 30B A3B Instruct perform on benchmarks?

Artificial Analysis reports an Intelligence Index of about 10, above the median of 6 for similar open-weight non-reasoning models, with 40.3% on LiveCodeBench and 15.2% on Terminal-Bench Hard. These are independent measurements: solid for routine coding, weaker on difficult agentic terminal tasks.

Is Qwen3 Coder 30B A3B Instruct open-weight, and what does the API cost?

Yes, the weights are released under Apache 2.0, with 30.5B total and 3.3B active parameters. API cost depends on the host, and hosted rates for this model vary between providers. For current AnyAPI.ai rates, use the AnyTokens pricing component on this page.

* Benchmark data source: Artificial Analysis artificialanalysis.ai