What are the strengths of this model?

Apache 2.0 license for commercial use Lightweight and fast Multilingual support (including Japanese) Active community

What are the weaknesses of this model?

Performance lags behind frontier models Japanese processing inferior to specialized models Limited API options outside Alibaba Cloud

What are the best use cases?

Development/research with open-source models Edge deployment requiring lightweight models Multilingual applications Cost-focused commercial services

AlibabaOpen Source

Qwen3.6-27B

Name: Qwen3.6-27B
Price: 0.35 USD
Author: Alibaba

Compare this model

The latest version of Alibaba's open-source model. With 27B parameters, it is lightweight yet achieves high performance, and is available for commercial use under the Apache 2.0 license. Multilingual support has been enhanced, and Japanese language processing capabilities have been significantly improved from the previous generation.

Share:X Hatena LINE

Parameters

27B

Context Window

128K

License

Apache 2.0

Release Date

2026-04-22

Japanese Language Capability

✅High-Quality JP

Multilingual model with strong Japanese language processing capabilities.

API Pricing

Input Price (per 1M tokens)

$0.35

Output Price (per 1M tokens)

$1.4

Billing Mode: standard

Strengths

・Apache 2.0 license for commercial use
・Lightweight and fast
・Multilingual support (including Japanese)
・Active community

Weaknesses

・Performance lags behind frontier models
・Japanese processing inferior to specialized models
・Limited API options outside Alibaba Cloud

Use Cases

・Development/research with open-source models
・Edge deployment requiring lightweight models
・Multilingual applications
・Cost-focused commercial services

Deep Analysis

SWE-bench Verified

77.2%

Surpasses its larger predecessor Qwen3.5-397B-A17B (76.2%)

AIME 2026

94.1%

Superior math reasoning

Input Price

$0.60 / 1M tokens

Competitive pricing via API providers

Context Window

262K (native) / 1M (ext.)

Extended via YaRN

Architecture

Dense 27B

Simple deployment, no MoE routing

License

Apache 2.0

Fully open-source, permissive for commercial use

Strengths

・Flagship-level agentic coding performance, surpassing much larger models on SWE-bench
・Exceptional mathematical reasoning (AIME 94.1%)
・Long-context capability (up to 1M tokens) and simple dense architecture for easier deployment

Weaknesses

・Performance may lag behind predecessor on some knowledge and instruction-following benchmarks
・Not the top performer in EU-language multilingual tasks
・Context window extension (YaRN) may require extra setup

Competitor Comparison

Model	Arena	SWE	GPQA	Price
Gemma 4 27B (Google)	N/A	43.2%	68.4%	$0.0 / $0.0*
Qwen3.5-27B (Alibaba)	N/A	72.4%	85.5%	$0.0 / $0.0*

Overview

Qwen3.6-27B is Alibaba's latest open-source dense model, released in April 2026. It focuses on delivering state-of-the-art performance in a practical, widely-deployable 27B-parameter package. Its standout achievement is surpassing the performance of Alibaba's own much larger previous-generation MoE flagship (397B total parameters) on critical agentic coding benchmarks, while being significantly simpler to host and run due to its dense architecture. Positioned as a cost-effective, high-performance workhorse, the model excels in software engineering, mathematical reasoning, and long-context tasks. It supports multimodal inputs (text, images, video) and features a massive native context window of 262K tokens, extendable to over 1 million. The permissive Apache 2.0 license and competitive API pricing make it highly attractive for developers and commercial applications. While its strengths are pronounced in technical domains, community evaluations suggest potential trade-offs in certain general knowledge tasks and instruction adherence compared to its direct predecessor. It is a top contender for coding assistants, technical problem-solving, and applications requiring long-context processing within the open-weight model tier.

Benchmarks & Performance

Qwen3.6-27B demonstrates exceptional performance in coding, reasoning, and math, often competing with models far beyond its parameter count. **Coding & Agentic Benchmarks:** | Benchmark | Qwen3.6-27B | Key Comparison | |---|---|---| | SWE-bench Verified | 77.2% | Beats Qwen3.5-397B-A17B (76.2%) | | Terminal-Bench 2.0 | 59.3% | Beats Qwen3.5-397B-A17B (52.5%) | | LiveCodeBench v6 | 83.9% | | | SkillsBench | 48.2% | Beats Qwen3.5-397B-A17B (30.0%) | **Reasoning & Knowledge:** | Benchmark | Qwen3.6-27B | Key Comparison | |---|---|---| | AIME 2026 | 94.1% | Significantly outperforms Gemma 4 27B (52.1%) | | GPQA Diamond | 87.8% | Competitive with larger models | | MMLU-Pro | 86.2% | Beats Gemma 4 27B (75.8%) | | C-Eval | 91.4% | | *Note: Some independent tests (e.g., by Kaitchup) found Qwen3.5-27B outperforming on GPQA Diamond, highlighting evaluation variance.*

Detailed Comparison

**vs. Google Gemma 4 27B:** - **Strengths:** Qwen3.6-27B has a massive lead in coding (SWE-bench: 77.2% vs 43.2%) and math (AIME: 94.1% vs 52.1%). It also supports video input and has a larger native context window. - **Weaknesses:** Gemma 4 is often cited as having stronger output quality in European languages (e.g., French, German) and a more conservative safety alignment. - **Pricing:** Both are open-weight. Gemma 4 has a custom license with commercial restrictions; Qwen3.6-27B uses permissive Apache 2.0. API pricing for Qwen via providers is very competitive (e.g., $0.60/$3.60 per M tokens). **vs. Qwen3.5-27B (Previous Generation):** - **Strengths:** Significant leap in agentic coding (SWE-bench: 77.2% vs 72.4%, Terminal-Bench: 59.3% vs 41.6%) and math reasoning. - **Weaknesses:** Some benchmarks show regression in knowledge recall (GPQA, HLE) and instruction following (IFBench), as noted in independent analyses. - **Context:** Qwen3.6-27B is the clear upgrade for developers focused on coding and math, but users prioritizing general knowledge may observe mixed results.

Community Feedback

Developer and researcher reception has been largely positive, particularly around the model's coding capabilities. The ability of a dense 27B model to outperform a 397B MoE model on core software engineering benchmarks has been highlighted as a major achievement and a practical win for deployment. The model's integration with popular coding agents like Claude Code, OpenClaw, and Qwen Code has been well-received, easing adoption. However, discussions have also surfaced regarding benchmark discrepancies and potential trade-offs. Independent tests from sources like Kaitchup and BenchLM noted the model's lower performance on some knowledge benchmarks compared to its predecessor, sparking debates about evaluation methodologies and the true cost of optimizing for agentic tasks. A notable community event was the public documentation of "abliteration" techniques applied to Qwen3.6-27B, which sparked discussion about open model safety, modification accessibility, and the responsibilities of startups deploying such modified models. Overall, the model is seen as a powerful and accessible tool, with the community actively exploring its strengths and limitations.

Use Cases

1. **AI-Powered Coding Assistants & IDE Integrations:** The model's top-tier performance on SWE-bench and compatibility with tools like Claude Code makes it ideal for powering next-generation code assistants that can understand repositories, debug complex issues, and generate code across multiple files. 2. **Technical Documentation & Codebase Analysis:** With its 262K+ token context window, it can ingest entire large codebases or lengthy technical manuals, enabling developers to ask natural language questions about system architecture, function dependencies, or to generate comprehensive documentation. 3. **STEM Education and Research:** Its exceptional mathematical reasoning (AIME 94.1%) and scientific knowledge make it a strong candidate for building educational tutors that can solve advanced problems step-by-step and explain complex concepts. 4. **Automated Software Engineering Workflows:** The model can drive agentic pipelines for tasks like repository migration, test case generation, code review, and implementing features from issue descriptions, reducing the manual effort on routine engineering tasks.

Latest News

April 2026: Qwen3.6-27B was officially released by Alibaba, making it available through the Alibaba Cloud Model Studio API, Hugging Face, and ModelScope under the Apache 2.0 license. Release announcements emphasized its agentic coding prowess and long-context capabilities. The model family also expanded to include Qwen3.6-Plus (API-only) and the Qwen3.6-35B-A3B MoE sibling, providing a range of options for different performance and cost trade-offs. May 2026: Community researchers published detailed "abliteration" experiments on the model, demonstrating relatively efficient methods to reduce its refusal behaviors. This sparked industry discussion about the balance between open model utility, safety, and the responsibilities of commercial deployers.

Positioned as a cost-effective, high-performance workhorse, the model excels in software engineering, mathematical reasoning, and long-context tasks. It supports multimodal inputs (text, images, video) and features a massive native context window of 262K tokens, extendable to over 1 million. The permissive Apache 2.0 license and competitive API pricing make it highly attractive for developers and commercial applications.

While its strengths are pronounced in technical domains, community evaluations suggest potential trade-offs in certain general knowledge tasks and instruction adherence compared to its direct predecessor. It is a top contender for coding assistants, technical problem-solving, and applications requiring long-context processing within the open-weight model tier.

Sources

Analysis generated: 2026-05-23