Qwen
•
Qwen3 Coder Next
•
Released 
February 2026

Qwen
Qwen3 Coder Next

80B-parameter MoE coding model activating only 3B parameters per token, built for agentic software engineering with a 262K context.

Modality:
Text
PDF
model ID
qwen/qwen3-coder-next

Output Speed *

83.85
tok/s

Intelligence Index *

0
/ 100

Context Window *

262144
tokens

Input price

0.72
Anytoken

Output price

4.5
Anytoken
Agentic Coding Performance at 3B Active Parameters Qwen3 Coder Next is an open-weight coding model from Alibaba's Qwen team, released in February 2026 as the compact member of the Qwen3-Coder line. It is an 80B-parameter mixture-of-experts model that activates 3B parameters per token, trained for long-horizon tool use and recovery from execution failures. It is text-in, text-out and non-thinking only. That makes it a fit for coding agents, CLI/IDE tools and repository-level code work, not for multimodal or deep-reasoning tasks. Access Qwen3 Coder Next via API

Performance

Over 70% on SWE-Bench Verified With Only 3B Active Parameters

Qwen reports over 70% on SWE-Bench Verified with the SWE-Agent scaffold. Its technical report lists 70.6%, 71.1% and 71.3% across SWE-Agent, MiniSWE-Agent and OpenHands. Only 3B of 80B parameters are active per token, and Qwen says its SWE-Bench Pro results are comparable to models with 10–20x more active parameters. That makes it a candidate for agent loops where many sequential calls matter more than peak single-answer quality. These are vendor-reported figures, so validate them on your own repositories and scaffold.

Benchmarks

Independent Speed, Latency and Verbosity Data for Qwen3 Coder Next

Artificial Analysis measures a median 76.9 output tokens per second and 1.15 s time to first token across providers. Speed is about average, but latency is low versus comparable open-weight non-reasoning models (median 2.06 s). Among the top three providers, speed ranges from 79.6 to 173.7 tokens per second, so host choice matters. The same source calls the model very verbose: 34M tokens generated on its Intelligence Index versus a 26M median. IFBench is 35.2%, so verify strict instruction adherence yourself.

Output Speed

*
83.85
tok/s

Intelligence Index

*
0
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
0
%

HLE *

Adherence to multi-step structured instructions.
0
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

Qwen3 Coder Next takes text and returns text, with a 262,144-token native context window. Maximum output depends on the host: Alibaba Cloud Model Studio lists 65,536 output tokens and 204,800 input tokens, while Amazon Bedrock lists 16K output. Plan around the lowest cap on your host, because long single-pass file generation can hit it long before the context limit. The model runs in non-thinking mode only, so there are no reasoning budgets to tune.
Verified Specifications — 
Qwen3 Coder Next
*
Input modalities
Text
PDF
output modalities
Text
Context window
262144
 tokens
Maximum output tokens
262144
Reasoning
No
Knowledge cutoff
February 2026
Pricing (standard)
0.72
 AnyTokens in
 / 
4.5
 AnyTokens out

Comparison

Qwen3 Coder Next or Qwen3 Coder 480B A35B for Coding Agents?

Qwen3 Coder 480B A35B was Qwen's previous flagship coder, with 480B total and 35B active parameters. Qwen3 Coder Next activates just 3B of 80B. Both are open-weight coding models aimed at agents. In Qwen's technical report the smaller model scores higher on Aider-Polyglot (66.2 vs 60.4) and gains on competitive-programming benchmarks. It trails slightly on FullStackBench-en (60.58 vs 62.54) and BIRD-SQL (63.56 vs 64.15). The practical question is whether the extra active compute buys anything for your workload.

Dimension
Qwen3 Coder Next
Qwen3 Coder 480B A35B Instruct
Context window *
262144
tokens
262144
tokens
Output speed *
83.85
tok/s
0.00
tok/s
Intelligence Index *
0
11.9
Input pricing
0.72
AnyToken
9
AnyToken
Output pricing
4.5
AnyToken
45
AnyToken
Knowledge cutoff *
February 2026
July 2025

Choose Qwen3 Coder Next when you run many sequential agent calls, want a much smaller model to self-host, or need strong multi-language editing and competitive-programming-style reasoning. Choose Qwen3 Coder 480B A35B when full-stack generation and text-to-SQL accuracy matter more than footprint, or when you want its documented extension to 1M tokens via YaRN. Both hosted IDs are on Alibaba Model Studio's October 10, 2026 retirement list, so verify availability on your host.

Limitations & Trade-offs

Where Qwen3 Coder Next falls short

1
Text-only, no thinking mode. The model accepts and returns text only, and Alibaba's documentation lists thinking mode as unsupported, so there is no reasoning budget to raise on hard problems. Screenshot-based UI debugging or image-grounded tasks need a multimodal model. Qwen's own summary says there is still much room for improvement on agent benchmarks, so route difficult multi-step reasoning to a dedicated reasoning model.
2
Output cap varies by host. The 262,144-token context window is not an output limit. Alibaba Cloud Model Studio lists 65,536 output tokens and 204,800 input tokens, while Amazon Bedrock lists 16K output. Workloads that generate whole modules or large migrations in one response can hit a lower host cap early. Check the cap of your serving route, or split generation into smaller agent steps.
3
High verbosity and variable speed. Artificial Analysis flags the model as very verbose, with 34M tokens generated on its Intelligence Index versus a 26M median. It also calls it expensive relative to open-weight non-reasoning peers of similar size. Speed varies widely by provider, from 79.6 to 173.7 tokens per second among the top three. High-volume agent loops should budget for long outputs and benchmark the exact host.
4
First-party endpoint retiring. Alibaba Model Studio lists qwen3-coder-next for retirement on October 10, 2026, with qwen3.7-plus as the recommended replacement. Aggregators note that this notice covers Model Studio hosting routes and does not establish retirement of other routes or the downloadable weights. Teams with direct Model Studio integrations should migrate or confirm an alternative host before relying on this ID.

Best-Fit Workloads

Where this model earns its place

01
Coding-agent backends Qwen reports over 70% on SWE-Bench Verified with SWE-Agent, and the model was trained for long-horizon tool use and recovery from failed executions. Qwen says it adapts to scaffolds such as Claude Code, Qwen Code, Cline and Kilo. It suits autonomous issue-fixing and test-repair loops. Because it has no thinking mode, pair it with a stronger planner for ambiguous tasks.
02
Repository-scale code analysis A 262,144-token native context window lets full modules, PR diffs and API specs fit in one prompt. That suits code review, dependency tracing and technical-debt audits. Alibaba Cloud lists a 204,800-token maximum input, so size inputs to your host. Responses can run long, so ask for concise findings.
03
Always-on, high-volume agent loops Only 3B of 80B parameters are active per token, and Qwen positions the model as cost-effective for agent deployment. That suits CI bots and background agents making many sequential calls. Artificial Analysis measures low median latency (1.15 s TTFT). The trade-off is verbose output, so cap response length in cost-sensitive pipelines.
04
Multi-language editing and text-to-SQL Qwen's technical report gives 66.2 on Aider-Polyglot, above the 480B coder's 60.4, plus 83.66 on Spider and 63.56 on BIRD-SQL. That supports cross-language code edits and SQL generation inside developer tools. Full-stack generation scores are slightly lower than the larger coder's, so test web-heavy tasks separately.
Pricing in anytokens via AnyAPI
Input
0.72
₳
Output
4.5
₳
Cache write
—
₳
Cache read
—
₳

Integration

Access Qwen3 Coder Next via AnyAPI.ai

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

01

One API integration

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

02

Easy model switching

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

03

Flexible for production

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

04

Multi-model applications

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

Frequently Asked Questions

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

Qwen3 Coder Next has a native context window of 262,144 tokens. Maximum output is separate and depends on the host: Alibaba Cloud Model Studio lists 65,536 output tokens and a 204,800-token maximum input, while Amazon Bedrock lists 16K output tokens. Size prompts and expected responses to your host's limits.

No to both. Qwen3 Coder Next accepts text input and returns text output. It runs only in non-thinking mode and does not emit think blocks, and Alibaba lists thinking mode as unsupported. For screenshot-based debugging or explicit reasoning budgets, use a multimodal or reasoning model.

Qwen reports over 70% on SWE-Bench Verified using the SWE-Agent scaffold. Its technical report lists 70.6% (SWE-Agent), 71.1% (MiniSWE-Agent) and 71.3% (OpenHands). These are vendor-reported results, and scores vary with scaffold and harness, so test the model on your own repositories before committing.

Yes. Qwen3 Coder Next is an open-weight model with 80B total and 3B active parameters, published under Apache 2.0 on Hugging Face. It can be self-hosted with engines such as vLLM, SGLang and llama.cpp. Hosted API access is also offered by providers including Amazon Bedrock, Together AI and Novita.

Alibaba Cloud Model Studio lists the qwen3-coder-next ID for retirement on October 10, 2026, recommending qwen3.7-plus as the replacement. That notice covers Model Studio hosting routes, and it does not establish retirement of other hosting routes or the open weights. Confirm current availability on your chosen API route.

* Benchmark data source: Artificial Analysis artificialanalysis.ai