Originally published on Remote OpenClaw.
Qwen is the largest and most complete open-source model family available in 2026. Alibaba's Qwen3 series spans 8 models from 0.6B to 235B parameters, all released under the Apache 2.0 license. The flagship Qwen3-235B-A22B scores 95.6 on ArenaHard, 77.1 on LiveBench, and leads on CodeForces Elo — putting it neck-and-neck with Gemini 2.5 Pro as the strongest open-weight generalist. Qwen3.5, released February 2026, extended the family to 397 billion parameters with 201 language support and throughput up to 19x faster than the previous generation.
This is the general Qwen model review covering Alibaba's AI strategy, the full model family, and competitive positioning. If you are looking for Qwen models specifically inside OpenClaw: read Best Qwen Models for OpenClaw, which covers DashScope configuration, Ollama setup, and model IDs for that workflow.
Key Takeaways
- Qwen3-235B-A22B is the flagship MoE model with 235B total / 22B active parameters, scoring 95.6 on ArenaHard and 85.7 on AIME'24 — competitive with the best closed-source models.
- Qwen3.5 (February 2026) scales to 397B total parameters with 17B active, supports 201 languages, and delivers 8.6x-19x throughput improvement over the previous generation.
- Qwen3-Coder 480B is the dedicated coding model with 480B total / 35B active parameters, trained on 7.5 trillion tokens (70% code), matching Claude Sonnet 4 on agentic coding benchmarks.
- The entire Qwen3 family is open-source under Apache 2.0, with dense models (0.6B to 32B) and MoE models (30B-A3B and 235B-A22B) that run on everything from phones to H100 clusters.
- DashScope API pricing starts at $0.01/M input tokens for the smallest model and $0.78/M for Qwen3 Max, with a free tier of 1M input + 1M output tokens for new accounts.
In this guide
- Alibaba's Open-Source AI Strategy
- The Qwen 3 Model Family
- Benchmark Comparison vs Llama and DeepSeek
- Open-Source Advantages and Apache 2.0
- Hosting Options and Pricing
- Limitations and Tradeoffs
- FAQ
Alibaba's Open-Source AI Strategy
Alibaba Cloud's Qwen team has pursued the most aggressive open-source AI strategy of any major tech company in the world. Every core Qwen model — including the flagship 235B-parameter MoE — is released under the Apache 2.0 license, which permits unrestricted commercial use, modification, and redistribution without royalty payments.
This stands in contrast to Meta's Llama approach (which uses a custom license with usage restrictions above certain thresholds), DeepSeek's selective open-sourcing, and of course the fully closed models from OpenAI and Anthropic.
The strategic logic is straightforward. Alibaba Cloud generates revenue from cloud compute and DashScope API access, not model licensing. Open-sourcing Qwen drives adoption, which drives API usage and cloud infrastructure revenue. As of April 2026, Qwen models are among the most downloaded model families on Hugging Face, with Qwen3 available on Hugging Face, GitHub, ModelScope, and Ollama.
The breadth of the lineup matters as much as the licensing. Qwen3 is not just a single frontier model — it is a family designed to cover every deployment tier from edge devices (0.6B, 1.7B) through laptops (4B, 8B, 9B) to data centers (32B, 235B, 480B). This full-stack approach is what gives Qwen a structural advantage over competitors that only release one or two model sizes.
The Qwen 3 Model Family
Qwen3 includes six dense models and two Mixture-of-Experts models, plus Qwen3.5 and Qwen3-Coder as major extensions released in 2026.
Model
Type
Total Params
Active Params
Context
License
Qwen3-0.6B
Dense
0.6B
0.6B
32K
Apache 2.0
Qwen3-1.7B
Dense
1.7B
1.7B
32K
Apache 2.0
Qwen3-4B
Dense
4B
4B
32K
Apache 2.0
Qwen3-8B
Dense
8B
8B
32K
Apache 2.0
Qwen3-14B
Dense
14B
14B
32K
Apache 2.0
Qwen3-32B
Dense
32B
32B
32K
Apache 2.0
Qwen3-30B-A3B
MoE
30B
3B
32K
Apache 2.0
Qwen3-235B-A22B
MoE
235B
22B
32K+
Apache 2.0
Qwen3.5 (Feb 2026)
MoE
397B
17B
262K
Apache 2.0
Qwen3-Coder 480B
MoE
480B
35B
262K (1M w/ YaRN)
Apache 2.0
Qwen3-235B-A22B is the original Qwen3 flagship, launched April 29, 2025. It introduced hybrid reasoning — the ability to switch between a "thinking mode" for complex multi-step tasks and a "non-thinking mode" for fast general responses. Trained on 36 trillion tokens (double Qwen2.5), it was the first open-weight model to match commercial frontier models on ArenaHard.
Qwen3.5, released February 16-17, 2026, is the latest generation. It scales to 397B total parameters with 17B active, uses 256 experts with 8 routed experts and 1 shared expert, and supports a native 262K token context window. The throughput gains are dramatic: 8.6x faster than Qwen3-Max at 32K context and 19x faster at 256K. Qwen3.5 also introduces Gated Delta Networks and expands language support to 201 languages and dialects.
Qwen3-Coder 480B-A35B is the dedicated coding model, released July 2025 with 480B total parameters and 35B active. It was trained on 7.5 trillion tokens with 70% code focus and achieves state-of-the-art results among open models on SWE-Bench Verified, matching Claude Sonnet 4 on agentic coding benchmarks.
The small models deserve attention too. Qwen3.5-9B beats OpenAI's gpt-oss-120B on several benchmarks while running on standard laptops. Qwen3-4B rivals Qwen2.5-7B performance at roughly half the memory footprint.
Key numbers to know
Marketplace
Free skills and AI personas for OpenClaw — browse the marketplace.
Benchmark Comparison vs Llama and DeepSeek
Qwen3-235B-A22B is the top-performing open-weight generalist model as of April 2026, with benchmark scores that put it ahead of both Llama 4 and DeepSeek V3.2 on most categories.
Benchmark
Qwen3-235B
Llama 4 Maverick
DeepSeek V3.2
ArenaHard
95.6
~90
~91
LiveBench
77.1
~72
~74
AIME'24
85.7
~78
89.3
HumanEval (Qwen3-32B)
88.0
—
82.6
GPQA Diamond
77.2
~69
~72
MMLU
~84
85.5
~83
CodeForces Elo
1
—
2
The competitive picture breaks down by category. Qwen3-235B leads on general intelligence (ArenaHard, LiveBench), reasoning (GPQA Diamond), and competitive programming (CodeForces). DeepSeek V3.2 is the strongest on pure mathematical reasoning with an AIME 2025 score of 89.3. Llama 4 Maverick posts the highest raw MMLU at 85.5 and has an unmatched 10-million-token context window through Scout.
On coding specifically, Qwen3-32B reports 88.0% on HumanEval — higher than DeepSeek V3.2 Speciale at 82.6%. Qwen3-Coder 480B extends this further with state-of-the-art open-model results on SWE-Bench Verified.
The cost comparison is less clear-cut. Qwen3-32B and Llama 4 Scout are essentially tied at $0.78-$0.83 per million tokens on comparable hardware (both run on a single H100). DeepSeek V3.2 Speciale is substantially more expensive at $13.33/M tokens and requires 8x H100 GPUs. For cost-per-intelligence, Qwen3 and Llama 4 are the strongest values in the open-weight space.
Open-Source Advantages and Apache 2.0
Qwen's Apache 2.0 licensing is the most permissive among the three dominant open-weight model families, and this matters for three practical reasons.
No usage thresholds. Meta's Llama 4 license includes restrictions for applications with more than 700 million monthly active users, which technically affects large platforms and creates legal ambiguity for enterprise teams. Apache 2.0 has no such limitations — commercial use, modification, and redistribution are unrestricted at any scale.
Self-hosting flexibility. Every Qwen3 model is downloadable from Hugging Face, GitHub, and ModelScope, and is available through Ollama for local deployment. The range of model sizes means you can self-host at every hardware tier — from Qwen3-0.6B on a Raspberry Pi to Qwen3-235B on a multi-GPU server.
Fine-tuning ecosystem. Apache 2.0 allows publishing fine-tuned derivatives without restrictions. This has created a large Qwen fine-tuning ecosystem on Hugging Face, with thousands of community-tuned variants for specific languages, domains, and use cases. The Qwen3.5-35B-A3B-Base model is also open-sourced alongside the instruct-tuned versions, giving researchers direct access to the base model for custom training.
The practical result is that Qwen has become the default starting point for teams building on open-weight models in 2026. Alibaba's strategy of full-stack open-sourcing — from tiny edge models to frontier-scale — gives the ecosystem a breadth that neither Meta nor DeepSeek matches.
Hosting Options and Pricing
Qwen models are available through Alibaba's own DashScope/Model Studio API, through third-party providers like OpenRouter, and through self-hosting via Ollama, vLLM, or other inference frameworks.
Model
DashScope Input (per 1M)
DashScope Output (per 1M)
OpenRouter Input (per 1M)
Qwen3 Max (235B)
$0.78
$3.90
~$0.83
Qwen3.5 Plus
~$1.04
~$5.20
~$1.10
Qwen3-32B
~$0.25
~$1.25
~$0.30
Qwen3.5-9B
~$0.05
~$0.25
~$0.08
Qwen3.5-0.8B
$0.01
$0.05
—
DashScope offers a free tier of 1 million input tokens + 1 million output tokens for new accounts, valid for 90 days. DashScope pricing is tiered based on the input size of each individual request (not cumulative session tokens), so costs vary depending on your prompt structure.
For local deployment, the smaller Qwen models are among the easiest to self-host. Qwen3.5-9B runs comfortably on consumer hardware with 8+ GB VRAM. Qwen3-32B fits on a single H100 or can run quantized on high-end consumer GPUs with 24+ GB VRAM. Qwen3-4B requires only ~3.5 GB RAM, making it viable on older laptops.
Qwen3-Coder 480B and Qwen3-235B require multi-GPU setups for full-precision inference, though quantized versions are available on Ollama. The throughput advantage of Qwen3.5 — up to 19x faster than the previous generation at 256K context — makes it significantly cheaper to self-host per-token than earlier Qwen generations.
Limitations and Tradeoffs
Qwen's breadth is a strength, but it creates its own set of tradeoffs.
Model selection complexity. With 10+ models across three generations (Qwen3, Qwen3.5, Qwen3-Coder), choosing the right one requires understanding the tradeoffs between dense vs MoE, base vs instruct, and generation differences. This is a real friction cost compared to model families that offer fewer, simpler choices.
Native context is shorter than competitors. Qwen3's native context is 32,768 tokens (extending to 131,072 with YaRN). Qwen3.5 extends this to 262,144 natively. But both trail Llama 4 Scout's 10M context and MiniMax-Text-01's 4M context by a wide margin. For extreme long-context use cases, Qwen is not the right family.
Mathematical reasoning trails DeepSeek. DeepSeek V3.2 holds an AIME 2025 score of 89.3 versus Qwen3-235B's 85.7. For math-heavy workloads, DeepSeek remains stronger.
DashScope API availability varies by region. The US (Virginia) deployment mode has no free quota and may have different latency characteristics than the primary Asian endpoints. Teams outside Asia should benchmark latency before committing to DashScope as a primary provider.
Qwen3-Coder 480B is not for local deployment. Despite being open-weight, a 480B MoE model is not something you run on consumer hardware. Self-hosting requires serious infrastructure. For most individual developers, the 32B dense model or the 30B-A3B MoE model is the practical ceiling for local use.
Related Guides
- Best Qwen Models for OpenClaw
- Best Llama Models in 2026
- Best DeepSeek Models in 2026
- Best Open-Source Models in 2026
FAQ
What is the best Qwen model in 2026?
Qwen3-235B-A22B is the best Qwen model for general use in 2026, scoring 95.6 on ArenaHard and leading on CodeForces Elo. For coding specifically, Qwen3-Coder 480B-A35B is the strongest, matching Claude Sonnet 4 on agentic coding benchmarks. For local deployment on consumer hardware, Qwen3.5-9B is the recommended starting point.
How does Qwen compare to Llama 4?
Qwen3-235B outperforms Llama 4 Maverick on ArenaHard (95.6 vs ~90), GPQA Diamond (77.2 vs ~69), and coding benchmarks. Llama 4 wins on raw MMLU (85.5 vs ~84) and has a vastly larger context window (10M through Scout). Both cost roughly the same per token when self-hosted on a single H100. Qwen's Apache 2.0 license is more permissive than Llama's custom license.
Is Qwen really open source?
Qwen3 models are released under the Apache 2.0 license, which allows unrestricted commercial use, modification, and redistribution. The weights, configuration files, and base models are all available on Hugging Face, GitHub, and ModelScope. This is the most permissive licensing among the three major open-weight model families (Qwen, Llama, DeepSeek).
How much does the Qwen API cost?
DashScope pricing for Qwen3 Max starts at $0.78 per million input tokens and $3.90 per million output tokens. The smallest model (Qwen3.5-0.8B) costs $0.01/M input. New accounts get a free tier of 1M input + 1M output tokens valid for 90 days.
Can I run Qwen models locally?
Qwen3.5-9B runs on consumer hardware with 8+ GB VRAM and is available through Ollama. Qwen3-32B fits on a single GPU with 24+ GB VRAM. Qwen3-4B requires only ~3.5 GB RAM and runs on older laptops. The larger MoE models (235B, 480B) require multi-GPU infrastructure for full-precision inference.
Top comments (0)