Qwen3's efficient coding model (30B MoE, 3B active). Fast code generation at low cost.
Qwen3-Coder-30B-A3B-Instruct is the efficient member of Alibaba Qwen's Qwen3-Coder family: a 30.5B-total / 3.3B-active Mixture-of-Experts coding model built for agentic coding, tool calls, and repository-scale context at much lower serving cost than the 480B flagship. The model card positions it as a streamlined coder that preserves strong performance on agentic coding, browser-use, and foundational coding tasks while improving throughput and deployability.
For TheRouter users, this is the practical high-throughput Qwen coder: choose it for Cursor-style assistants, Cline, code review prefilters, test generation, and batch coding workflows where the 480B model would be overkill. It keeps the same 262K native context headline and non-thinking output behavior, so production clients do not need to strip <think> blocks from responses.
| Type | Rate |
|---|---|
| Input | $0.486 / 1M tokens |
| Output | $2.43 / 1M tokens |
| Architecture | Mixture-of-Experts causal language modelhuggingface.co ↗ | verified |
| Parameters | 30.5B total / 3.3B activatedhuggingface.co ↗ | verified |
| Layers | 48huggingface.co ↗ | verified |
| Experts | 128 experts; 8 activated per tokenhuggingface.co ↗ | verified |
| Native context | 262,144 tokens; extendable to 1M with YaRN in self-hosted deploymentshuggingface.co ↗ | verified |
| Output mode | Non-thinking only; does not generate <think> blockshuggingface.co ↗ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
Agentic Coding / Browser-Use / Foundational Coding The 30B model card states significant performance among open models on agentic coding, agentic browser-use, and foundational coding tasks, but does not publish exact scores in the fetched text. | Strong open-model performance; exact 30B-specific score not exposed in fetched model-card text | huggingface.co/Qwen ↗ | |
SWE-Bench Verified Qwen's family announcement discusses SWE-Bench Verified for Qwen3-Coder and the flagship 480B model; do not infer a 30B score until Alibaba publishes one. | — | Not publicly disclosed | — |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "qwen/qwen3-coder-30b",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Use the standard OpenAI-compatible chat completions endpoint for code edits, tests, and explanations. Keep prompts concrete: include the target file, failing assertion, or API contract.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen/qwen3-coder-30b",
"messages": [
{"role": "system", "content": "You are a senior software engineer. Return concise patches and tests."},
{"role": "user", "content": "Write unit tests for this parser and cover edge cases: ..."}
],
"temperature": 0.7,
"max_tokens": 4096
}'Qwen introduced a streamlined 30.5B-total / 3.3B-active coder that keeps the family emphasis on long context, agentic coding, browser-use, and tool-use while lowering deployment and routing cost.
The initial Qwen3-Coder announcement introduced the family as an agentic coding line with Qwen Code, OpenAI-compatible API examples, Cline setup notes, and Claude Code-style integration paths.
The 30B model is the efficient sibling: 30.5B total and 3.3B active parameters versus 480B total and 35B active for the flagship. Use 30B for high-throughput coding, tests, summaries, and cost-sensitive agents; reserve 480B for higher-stakes autonomous edits or harder repository reasoning.
Yes. The model card highlights agentic coding and includes an OpenAI-compatible tool-calling example. Through TheRouter, use the normal tools and tool_choice fields on /v1/chat/completions.
No. Qwen states this model supports only non-thinking mode and does not generate <think></think> blocks. That makes it easier to use in production code clients that expect direct output.
Yes. TheRouter exposes the model through the OpenAI-compatible chat endpoint, so clients that let you set baseURL, API key, and model name can route to qwen/qwen3-coder-30b.
The curated model card records the model's native context claim, while operational context and max completion values still come from standard-models.yaml and the selected upstream route. Validate your largest prompts before production rollout.
| source | URL | retrieved | |
|---|---|---|---|
| Architecture | huggingface.co ↗ | 2026-06-03 | verified |
| Parameters | huggingface.co ↗ | 2026-06-03 | verified |
| Layers | huggingface.co ↗ | 2026-06-03 | verified |
| Experts | huggingface.co ↗ | 2026-06-03 | verified |
| Native context | huggingface.co ↗ | 2026-06-03 | verified |
| Output mode | huggingface.co ↗ | 2026-06-03 | verified |
| Agentic Coding / Browser-Use / Foundational Coding | huggingface.co/Qwen ↗ | 2026-06-03 | verified |
| SWE-Bench Verified | qwenlm.github.io ↗ | 2026-06-03 | unknown |
| Qwen releases Qwen3-Coder-30B-A3B-Instruct | huggingface.co/Qwen ↗ | 2026-06-03 | verified |
| Qwen launches the Qwen3-Coder family and Qwen Code | qwenlm.github.io ↗ | 2026-06-03 | verified |
| How is qwen/qwen3-coder-30b different from qwen/qwen3-coder-480b? | huggingface.co ↗ | 2026-06-03 | to verify |
| Does Qwen3-Coder-30B support tool calling? | huggingface.co ↗ | 2026-06-03 | to verify |
| Does it return hidden reasoning or <think> blocks? | huggingface.co ↗ | 2026-06-03 | to verify |
| Can I use it from Cursor, Cline, or another OpenAI-compatible client? | qwenlm.github.io ↗ | 2026-06-03 | to verify |
| Is the 262K context a TheRouter guarantee? | huggingface.co ↗ | 2026-06-03 | to verify |