On April 28, 2025, the Alibaba Qwen Team dropped eight model weights simultaneously on Hugging Face, all under Apache 2.0. No press embargo. No staged rollout. Just a technical blog post, a paper, and download links ranging from a 0.6-billion-parameter model that runs on a smartphone to a 235-billion-parameter Mixture-of-Experts model that outperforms DeepSeek R1 on reasoning benchmarks.
The release had been building for two years. The Qwen team at Alibaba’s DAMO Academy had been shipping model generations with unusual regularity — Qwen 1.0 in September 2023, Qwen 1.5 in February 2024, Qwen 2.0 in June 2024, Qwen 2.5 in September 2024, Qwen 2.5-Max (MoE) in January 2025 — each a meaningful step forward. Qwen 3 was not an incremental update. It brought a new architecture feature — hybrid thinking mode — to all eight sizes simultaneously, while also pushing the top-of-range MoE to genuine frontier performance on public benchmarks.
The result is the most capable open-weight model family as of mid-2025: matching or exceeding DeepSeek R1, competitive with GPT-4.1 and Claude 3.7 Sonnet, fully open-licensed, and deployable at any scale from edge devices to data centers. This review covers Qwen 3 in detail.
Who Is the Qwen Team?
Alibaba is China’s largest e-commerce company and one of its largest cloud providers. Alibaba Cloud (阿里云, Aliyun) is roughly equivalent to AWS in China’s domestic market — dominant in cloud infrastructure, with significant international presence across Southeast Asia, the Middle East, and Europe. The company is publicly traded on the New York Stock Exchange and the Hong Kong Stock Exchange.
The Qwen Team operates within DAMO Academy (达摩院, Discovery, Adventure, Momentum and Outlook), Alibaba’s research organization founded in 2017 with an initial $15 billion commitment over three years. DAMO Academy has published research across AI, quantum computing, fintech, and human-computer interaction. It is not a startup or spin-off — it is a well-resourced corporate research lab with access to Alibaba’s cloud infrastructure and business use cases.
Alibaba has strong competitive motivation to invest in frontier AI. Its domestic rivals — ByteDance (Doubao), Baidu (ERNIE Bot), and Tencent (Hunyuan) — all have competing model families. Internationally, Alibaba Cloud’s ability to offer competitive AI capabilities matters for its positioning against AWS Bedrock, Google Cloud Vertex AI, and Azure OpenAI Service. The Qwen models are both a research project and a commercial product line.
The team’s track record across the Qwen 1.x and 2.x series established credibility. Qwen 2.5-72B (September 2024) was widely recognized as the strongest open-weight 72B model at the time, competitive with models several times larger. By the time Qwen 3 launched, the team had two years of model training infrastructure, dataset curation, and benchmark evaluation accumulated.
The Qwen History
Qwen 1.0 (September 2023): 7B, 14B, and 72B parameter dense models. Early strong performer, introduced with comprehensive documentation and full weights on Hugging Face. The 7B model was competitive with LLaMA 2-70B at a fraction of the size.
Qwen 1.5 (February 2024): Added 0.5B, 1.8B, 4B, 32B sizes for broader deployment range. Extended context to 32K tokens for some sizes. Strong multilingual improvement. Introduced Qwen1.5-MoE-A2.7B — an early MoE experiment at the small scale.
Qwen 2.0 (June 2024): Major step forward. Qwen2-7B matched or exceeded many larger models. Qwen2-72B was the best open-weight model of its generation on most benchmarks. Apache 2.0 license extended across the full lineup. 128K context for the 72B model.
Qwen 2.5 (September 2024): Introduced specialized variants — Qwen2.5-Coder (code-focused, up to 72B), Qwen2.5-Math (mathematics, up to 72B), plus the base Qwen2.5 series. The coding and math variants became widely used in academic and professional settings. The 72B model pushed closer to frontier performance on math and coding benchmarks.
Qwen 2.5-Max (January 2025): First large-scale Qwen MoE, deployed on Alibaba Cloud’s Qwen AI platform. Not publicly released as weights — API-only at launch. A preview of the MoE architecture direction that would be fully released with Qwen 3.
Qwen 3 (April 28, 2025): Full model family refresh with hybrid thinking mode, new MoE models, improved instruction following, and 36 trillion token pretraining for the largest variants.
The Qwen 3 Model Lineup
Qwen 3 released eight distinct model sizes simultaneously:
Dense Models
| Model | Parameters | Context | Active Params |
|---|---|---|---|
| Qwen3-0.6B | 0.6B | 32K | 0.6B |
| Qwen3-1.7B | 1.7B | 32K | 1.7B |
| Qwen3-4B | 4.0B | 128K | 4.0B |
| Qwen3-8B | 8.0B | 128K | 8.0B |
| Qwen3-14B | 14.0B | 128K | 14.0B |
| Qwen3-32B | 32.0B | 128K | 32.0B |
Mixture-of-Experts Models
| Model | Total Params | Active Params | Context |
|---|---|---|---|
| Qwen3-30B-A3B | 30.0B | 3.0B | 128K |
| Qwen3-235B-A22B | 235.0B | 22.0B | 128K |
The naming convention follows [Total]-A[Active] — so Qwen3-235B-A22B activates 22 billion parameters per forward pass from a total 235 billion, and Qwen3-30B-A3B activates 3 billion from 30 billion. Both MoE models support the full 128K context window.
All eight models support hybrid thinking mode — the same model can run in either thinking or non-thinking mode, making this the first full-lineup deployment of the architectural pattern.
Architecture
Hybrid Thinking Mode
The central innovation of Qwen 3 is not the MoE scaling or the benchmark numbers — it is the hybrid thinking mode that works across all eight sizes.
In thinking mode, the model generates an extended internal reasoning sequence enclosed in <think> / </think> tags before producing a final answer. This reasoning trace is visible to the user (or developer) and constitutes the model’s deliberation process: exploring multiple solution paths, checking intermediate steps, identifying errors, and revising conclusions. The final answer follows the reasoning block and is produced after that deliberation is complete.
In non-thinking mode, the model responds directly without the reasoning trace — faster, more concise, and appropriate for conversational or simple factual queries where extended deliberation adds latency without value.
The key feature: both modes run the same model weights. There is no separate “thinking model” and “fast model” — one model instance, one set of parameters, two operational modes switchable per request. This is a significant engineering and UX advantage over systems that require users to choose between a reasoning model and a faster chat model at setup time.
Technically, thinking mode is enabled by including a specific thinking-triggering token at the start of the generation. The model was trained with a mixture of thinking and non-thinking examples, allowing it to fluently transition between both modes. The enable_thinking=True parameter in the Qwen3 inference API controls this at the API layer.
This is architecturally similar to the approach Anthropic uses for Claude’s extended thinking (available in Claude 3.7 Sonnet and Claude 4), though Anthropic’s implementation uses a separate budget token mechanism. It differs from DeepSeek R1’s approach, where R1 and its distilled variants always think — there is no non-thinking mode built into R1.
The practical result: Qwen 3 models can serve both as fast chat models and as reasoning engines, from the same deployment. For developers building applications that sometimes need quick answers and sometimes need careful multi-step reasoning, this eliminates the need to route requests to different model endpoints.
Transformer Architecture
Qwen 3 uses a standard transformer decoder with several modern modifications:
Grouped Query Attention (GQA): Rather than having one key/value head per query head (standard multi-head attention), GQA groups multiple query heads to share a single key/value head. This reduces KV cache memory at inference without meaningful quality degradation — an important property for long-context and batch inference.
RoPE (Rotary Position Embeddings): RoPE encodes position information relative to token positions rather than absolute positions, enabling better generalization to sequence lengths not seen during training. All Qwen 3 models use RoPE.
SwiGLU activation: A Swish-Gated Linear Unit activation function in the FFN layers. This is standard in modern high-quality transformers (used in LLaMA, Gemma, Mistral, etc.).
RMSNorm: Root Mean Square Layer Normalization, computationally cheaper than LayerNorm with equivalent quality.
No sliding window attention: Unlike some models that use local attention patterns to manage context, Qwen 3 uses full causal self-attention across the context window. This is a quality-over-efficiency trade-off — full attention is more expensive but captures all cross-token relationships.
MoE Architecture (Qwen3-235B-A22B and Qwen3-30B-A3B)
The MoE architecture follows the pattern established by DeepSeek’s work, though with Alibaba’s own implementation:
Expert routing: Each token is routed to a subset of experts in each MoE layer. The routing is learned end-to-end. The model learns which experts to activate for which types of tokens.
Expert count and active ratio: Qwen3-235B-A22B activates 22 billion of 235 billion parameters per token — roughly a 9.4% activation rate. Qwen3-30B-A3B activates 10% of total parameters. These ratios are similar to DeepSeek V3 (37B/671B = 5.5%) and Mixtral (12.9B/47B = 27.5%), placing Qwen 3’s MoE in the middle range of activation density.
Training efficiency: MoE models train more efficiently per unit of compute than dense models of equivalent total parameter count, because each training example activates only a fraction of the model. The Qwen3-235B-A22B achieves performance comparable to much larger dense models while requiring roughly the compute of a ~22B dense forward pass.
Training Data and Scale
Qwen 3’s largest models were pretrained on approximately 36 trillion tokens — more than double the Qwen 2.5 series and nearly 2.5x DeepSeek V3’s 14.8 trillion tokens.
The training data spans 100+ languages, with a deliberate emphasis on STEM content, code, and multilingual coverage. Alibaba disclosed that the data pipeline included extensive deduplication, quality filtering, and synthetic data generation for underrepresented domains. The multilingual depth is reflected in the benchmark results: Qwen 3 models perform strongly on Chinese, Japanese, Korean, Arabic, and other non-English evaluations where many Western labs have thinner coverage.
Post-training followed a multi-stage process:
- Supervised Fine-Tuning (SFT): Standard instruction-following training on curated examples
- Long Chain-of-Thought Cold Start: SFT on extended thinking traces to initialize the thinking mode behavior
- Reasoning RL: Reinforcement learning to improve thinking mode quality, using accuracy-based rewards for math and code
- Thinking/Non-Thinking Fusion: Training to allow seamless switching between modes within the same model
This staged process is broadly similar to DeepSeek R1’s approach (cold-start SFT → RL → alignment), with Qwen 3 adding the hybrid mode fusion as an additional stage.
Benchmarks
Flagship: Qwen3-235B-A22B (Thinking Mode)
Qwen3-235B-A22B with thinking mode enabled establishes the performance ceiling for the Qwen 3 family:
| Benchmark | Qwen3-235B-A22B (Thinking) | DeepSeek R1 | GPT-4.1 | Claude 3.7 Sonnet |
|---|---|---|---|---|
| AIME 2024 | 85.7% | 79.8% | ~72% | 80.0% |
| MATH-500 | 97.4% | 97.3% | 95.2% | 96.2% |
| GPQA Diamond | 71.1% | 71.5% | ~66% | 70.0% |
| LiveCodeBench | 70.7% | 65.9% | 54.6% | ~55% |
| Codeforces ELO | ~2100 | 2029 | ~1800 | ~1800 |
AIME 2024 (American Invitational Mathematics Examination): 85.7% for Qwen3-235B-A22B vs. 79.8% for DeepSeek R1. This is the most cited comparison: Qwen 3 surpasses R1 on the benchmark that made R1 famous.
MATH-500: 97.4% vs. R1’s 97.3% — essentially tied. Both models effectively solve nearly all MATH-500 problems with thinking enabled.
LiveCodeBench: 70.7% for Qwen3-235B-A22B vs. R1’s 65.9% — meaningful advantage in competitive programming, Qwen 3 ahead.
GPQA Diamond: 71.1% vs. R1’s 71.5% — essentially equivalent on graduate-level science questions.
The pattern: Qwen3-235B-A22B in thinking mode is the best-performing open-weight model on public benchmarks as of April 2025. The only closed models that clearly exceed it are GPT-4.5 (if available) and Claude 3.7 Sonnet extended thinking on some tasks — and even these advantages are task-specific.
Flagship: Qwen3-235B-A22B (Non-Thinking Mode)
Non-thinking mode trades reasoning depth for speed:
| Benchmark | Qwen3-235B-A22B (Non-Thinking) | Qwen3-235B-A22B (Thinking) |
|---|---|---|
| AIME 2024 | ~52% | 85.7% |
| MATH-500 | 93.4% | 97.4% |
| GPQA Diamond | ~62% | 71.1% |
| LiveCodeBench | ~58% | 70.7% |
Non-thinking mode is roughly comparable to DeepSeek V3 (without R1’s reasoning): still very strong for general tasks, code, and factual queries, but the reasoning advantage disappears on competition-level math and logic problems.
Dense Model Performance
| Model | MMLU | MATH-500 (Think) | LiveCodeBench (Think) |
|---|---|---|---|
| Qwen3-32B | ~82% | 95.2% | ~65% |
| Qwen3-14B | ~79% | 93.1% | ~60% |
| Qwen3-8B | ~75% | 89.0% | ~53% |
| Qwen3-4B | ~69% | 82.6% | ~45% |
| Qwen3-1.7B | ~60% | ~70% | ~30% |
| Qwen3-0.6B | ~50% | ~50% | ~18% |
The smaller models have thinking mode too, though the reasoning quality is naturally lower. Qwen3-32B in thinking mode is competitive with DeepSeek R1-Distill-Qwen-32B (a distilled version of R1 that is not a full thinking model but uses R1’s reasoning style). Qwen3-8B is competitive with Llama 4 Scout (17B active parameters from 109B total) on math benchmarks despite having far fewer parameters.
MoE Efficiency: Qwen3-30B-A3B
| Model | Active Params | MATH-500 (Think) | LiveCodeBench |
|---|---|---|---|
| Qwen3-30B-A3B | 3B | ~90% | ~55% |
| Qwen3-4B (dense) | 4B | 82.6% | ~45% |
| Qwen3-8B (dense) | 8B | 89.0% | ~53% |
Qwen3-30B-A3B achieves performance between the 4B and 8B dense models while running at the inference cost of a 3B model. For deployment environments where per-token compute cost matters (edge servers, low-latency applications, high-throughput inference), this is a compelling efficiency point.
License: Apache 2.0
All eight Qwen 3 models are released under the Apache 2.0 license.
Apache 2.0 means:
- Commercial use: Fully permitted, no revenue cap, no user count limit
- Attribution: Required only in source code notices, not in product marketing or branding
- Derivative works: Permitted — you can fine-tune, distill, and redistribute derivatives
- Patent grant: Included — contributors grant patent rights
- No viral provisions: Apache 2.0 is not copyleft — derivatives do not need to be open-sourced
This is the most permissive practical license for open-weight AI models. For comparison:
| Model | License | Key Restrictions |
|---|---|---|
| Qwen 3 | Apache 2.0 | None meaningful |
| DeepSeek V3/R1 | MIT | None meaningful |
| Meta Llama 4 | Llama 4 Community License | Attribution + 700M MAU cap |
| Mistral | Apache 2.0 | None meaningful |
| GPT-4.1 | Proprietary | API-only, no weights |
| Claude 3.7/4 | Proprietary | API-only, no weights |
Both Qwen 3 and DeepSeek (MIT) offer essentially identical practical licensing freedom. Meta Llama 4’s community license restricts services with more than 700 million monthly active users — relevant only for the largest platforms, but still a deviation from full permissiveness. OpenAI and Anthropic release no weights at all.
The Apache 2.0 choice is deliberate strategy. Alibaba competes with AWS and Azure for cloud workloads internationally. Making Qwen 3 maximally usable — including for companies that need to deploy on other clouds or on-premises without usage caps or attribution requirements — expands the addressable market. The model itself is the product; the license is designed to maximize adoption, not protect a proprietary moat.
Multilingual Capability
Qwen 3’s multilingual training is a differentiated capability worth separate attention.
The training data covers 100+ languages, with particular depth in:
- Chinese (Simplified and Traditional) — native strength; Alibaba’s primary market
- Japanese — strong academic and enterprise user base
- Korean — competitive business environment in Northeast Asia
- Arabic — significant Middle Eastern cloud market
- Spanish, French, Portuguese, German — standard European coverage
- Southeast Asian languages (Thai, Vietnamese, Indonesian) — Alibaba Cloud’s expansion markets
The result is meaningful performance parity between English and Chinese on reasoning-heavy benchmarks — unusual for models trained primarily on English corpora. For non-English enterprise deployments, this makes Qwen 3 more useful than equivalently-rated English-native models.
For multilingual deployment specifically, Qwen 3 likely leads the open-weight field as of mid-2025.
Deployment and Hardware Requirements
Qwen3-235B-A22B
The flagship MoE requires substantial infrastructure. At 22B active parameters and 235B total:
- Full BF16 inference: approximately 470 GB GPU memory (4× H100-80GB minimum with tensor parallelism)
- INT4 quantized: approximately 120 GB GPU memory (2× H100-80GB possible)
- AWQ/GPTQ quantization: further reduction, some quality loss
This is comparable to DeepSeek V3 in hardware requirements. Not a model for consumer hardware.
Qwen3-30B-A3B
At 30B total / 3B active:
- Full BF16: approximately 60 GB (1× H100-80GB or 2× A100-40GB)
- INT4/AWQ: approximately 18 GB (1× 24GB consumer GPU, e.g., RTX 4090)
This is the “power user at home” sweet spot: near-32B quality at 3B inference cost.
Qwen3-14B and Qwen3-8B
- Qwen3-14B BF16: ~28 GB (1× A100-40GB or 2× 24GB consumer GPUs)
- Qwen3-8B BF16: ~16 GB (1× RTX 4090 or A10G)
Both run comfortably on high-end consumer hardware in BF16. These are the practical choice for local deployment and fine-tuning on a single consumer GPU.
Qwen3-4B and smaller
- Qwen3-4B: ~8 GB (high-end gaming GPU or laptop GPU)
- Qwen3-1.7B: ~3.5 GB (mid-range GPU)
- Qwen3-0.6B: ~1.2 GB (integrated graphics, mobile, Raspberry Pi 5)
The smallest models run on smartphones. With MLX on Apple Silicon, Qwen3-4B achieves real-time generation speeds on an M2 MacBook Pro.
Inference Framework Support
At launch, Qwen 3 had immediate support from:
- vLLM: Full support with tensor parallelism for large models
- Ollama: All sizes available via
ollama pull qwen3:xxxb - llama.cpp: GGUF quantization support from day one
- LM Studio: GUI access for local deployment
- MLX (Apple Silicon): Optimized inference for Mac
- Transformers (Hugging Face): Reference implementation
The breadth of day-one framework support reflects the Qwen 2.x series’ popularity — framework maintainers already had Qwen support code and updated it for Qwen 3.
Qwen 3 vs. The Competition
vs. DeepSeek R1
DeepSeek R1 was the previous frontier-performance open-weight reasoning model. Qwen3-235B-A22B with thinking enabled outperforms R1 on AIME 2024 (85.7% vs. 79.8%), ties on GPQA Diamond, and leads on LiveCodeBench.
Key differences:
- License: Both MIT/Apache 2.0 — equivalent permissiveness
- Thinking mode: Qwen 3 has hybrid (switchable); R1 always thinks
- Context: Both 128K
- Hardware: Qwen3-235B-A22B requires slightly less memory at full precision (22B active vs. 37B active for V3)
- Jurisdiction: Both PRC — same geopolitical risk profile
- Multilingual: Qwen 3 edges ahead on non-English tasks
For pure reasoning quality: Qwen3-235B-A22B > DeepSeek R1. For deployment flexibility (always-on vs. switchable thinking): Qwen 3 wins. For smaller model sizes: Qwen 3 has a much fuller lineup (0.6B–32B dense vs. R1’s distilled 1.5B–70B).
vs. Meta Llama 4 (Scout and Maverick)
| Dimension | Qwen3-235B-A22B | Llama 4 Maverick |
|---|---|---|
| GPQA Diamond | 71.1% | 69.8% |
| LiveCodeBench | 70.7% | 43.4% |
| License | Apache 2.0 | Llama 4 Community (700M MAU cap) |
| Reasoning mode | Hybrid thinking | None |
| Context | 128K | 1M |
| Multimodal | No (text-only) | Yes (early fusion) |
| Audio | No | No |
Qwen3-235B-A22B leads on coding and reasoning. Llama 4 Maverick leads on long context (1M vs. 128K) and multimodality (early fusion vision). For text-only reasoning workflows, Qwen 3 is stronger. For multi-modal or ultra-long-context tasks, Llama 4 has advantages.
vs. GPT-4.1
GPT-4.1 is API-only with no public weights. Qwen3-235B-A22B with thinking enabled matches or exceeds GPT-4.1 on most public benchmarks. For organizations that need on-premises deployment, full weight access, or cost control below GPT-4.1’s pricing ($2/$8 per million tokens), Qwen 3 is compelling. For organizations that need OpenAI’s reliability, support contracts, and managed service, GPT-4.1 is still the operational choice.
vs. Claude 3.7 Sonnet
Claude 3.7 Sonnet (with extended thinking) and Qwen3-235B-A22B are close competitors on benchmark metrics. Both have hybrid thinking architectures. Claude 3.7 has a larger context window (200K vs. 128K) and is API-only (no self-hosting). Qwen 3 is fully open-weight and license-free. For users who need Anthropic’s safety guarantees, alignment approach, and managed service, Claude 3.7 is the choice. For users who need open weights, Qwen 3 is the alternative most competitive with Claude.
Limitations
PRC Jurisdiction and Censorship
Like DeepSeek, Qwen is developed by a PRC-based organization. Unlike DeepSeek, which is a private hedge fund spinoff, Qwen comes from Alibaba — a publicly traded company with extensive ties to PRC regulatory structures and a history of responding to government pressure.
The models exhibit censorship on politically sensitive topics relevant to PRC governance: Tiananmen Square, Taiwan independence, Xinjiang, Tibet, and Hong Kong. The censorship is less obvious in the weights themselves (which can be fine-tuned to remove filters) but is present in the default instruction-tuned models. For enterprise deployments in regulated industries or government contexts where this matters, Qwen 3’s origin is a liability.
There is also geopolitical risk for US-based enterprises in regulated sectors. Defense contractors, intelligence-adjacent firms, and some financial institutions may face policy or compliance barriers to using Alibaba-developed model weights, even self-hosted.
No Native Multimodality in Qwen 3 Text Models
Qwen 3 (the text models covered in this review) are text-only. Alibaba has separate Qwen-VL (vision-language) and Qwen-Audio model families, but these are distinct products — not the same as GPT-4o or Llama 4’s native early-fusion multimodality where vision and text processing share the same transformer layers.
For applications that require image understanding, document analysis with visual elements, or video comprehension, the Qwen 3 text models require a separate vision model in the pipeline.
128K Context vs. Competitors
128K tokens is substantial for most tasks, but Llama 4 Maverick’s 1M token context and Claude 3.7 Sonnet’s 200K context are longer. For applications involving very long documents, entire codebases, or extended conversation history, the 128K limit is a real constraint. The thinking trace also consumes context — complex reasoning problems can use thousands of tokens for the <think> block, reducing effective available context for the actual query.
MoE Infrastructure Requirements
Qwen3-235B-A22B requires multi-GPU infrastructure even with quantization. Organizations without H100-class GPUs cannot run the flagship model. The 30B-A3B MoE is more accessible, but even it requires a well-configured 24GB GPU. For truly resource-constrained environments, the dense 8B or 14B models are the practical options, trading reasoning quality for deployability.
Production Availability
Qwen 3 models are available through:
Hugging Face: All 8 models, Apache 2.0, direct weight download. The standard path for self-hosting and fine-tuning.
Alibaba Cloud Qwen AI Platform (qwen.aliyun.com): API access with pay-per-token pricing. No weights needed — managed service option for organizations that don’t need on-premises deployment.
Together AI: Inference hosting for Qwen 3 models, competitive pricing, US-based infrastructure — the practical choice for US-based enterprises that want Qwen 3 capability without running on Alibaba Cloud.
Groq LPU: Qwen 3 smaller models available on Groq, near-zero latency inference.
Replicate, Fireworks AI, Anyscale: Third-party inference for Qwen 3 models.
Ollama: ollama pull qwen3:0.6b through ollama pull qwen3:235b-a22b — local deployment on consumer hardware.
Who Should Use Qwen 3?
Use Qwen3-235B-A22B (with thinking) if:
- You need the best available open-weight reasoning performance
- You need on-premises deployment with no usage caps
- You have H100-class infrastructure or access to third-party inference
- Your use case is math, code, science, or complex analysis
- Multilingual deployment (especially CJK languages) is a priority
Use Qwen3-30B-A3B if:
- You need near-32B quality at 3B inference cost
- Consumer GPU deployment (single RTX 4090 in INT4)
- Cost-sensitive high-throughput inference
Use Qwen3-14B or Qwen3-8B if:
- Local development and experimentation
- Fine-tuning on a single GPU
- Production deployments where hardware is A10G-class
Use Qwen3-4B or smaller if:
- Edge deployment (IoT, mobile, embedded systems)
- Extremely latency-sensitive applications
- Integration testing with minimal hardware
Do not use Qwen 3 if:
- Your organization has PRC-vendor restrictions (defense, intelligence, regulated finance)
- You need native multimodality — use Qwen-VL or a different model family
- You need >128K context — use Llama 4 Maverick or Claude 3.7 Sonnet
- You need Anthropic’s safety alignment methodology — use Claude
Rating: 4.5/5
Qwen 3 earns 4.5/5 — the highest rating we’ve given any open-weight model, tied with DeepSeek V3/R1.
Why 4.5:
- Frontier-class reasoning performance on public benchmarks, matching or beating DeepSeek R1
- Hybrid thinking mode that works across all eight sizes, including models small enough to run on edge devices
- Apache 2.0 license — the most permissive practical license in the open-weight space
- Eight sizes covering 0.6B to 235B, all released simultaneously, all with full weights and immediate framework support
- Exceptional multilingual performance (100+ languages with genuine CJK depth)
- Day-one support from vLLM, Ollama, llama.cpp, MLX — the entire inference ecosystem was ready
Why not 5:
- PRC jurisdiction creates geopolitical risk for some enterprise and government deployments
- No native multimodality (text-only; separate VL models needed for vision tasks)
- 128K context is good but shorter than Llama 4 (1M) or Claude 3.7 (200K)
- Flagship model requires multi-GPU infrastructure — not accessible for most individual developers without cloud access
The open-weight AI landscape as of May 2025 has two clear leaders: DeepSeek V3/R1 and Qwen 3. Both are Apache 2.0 / MIT. Both are Chinese. Both exceed most closed frontier models in at least some benchmark categories. For reasoning specifically, Qwen3-235B-A22B now leads. For the full size lineup (0.6B through 235B, all with thinking), Qwen 3 has no peer.
If you are evaluating open-weight models for any reasoning-intensive application, Qwen 3 belongs at the top of your list.
Next generation: Alibaba released Qwen 3.5 in February–March 2026, introducing a hybrid linear+sparse attention architecture (Gated DeltaNet), native early-fusion multimodal training, 262K context, 201 languages, and nine model sizes up to 397B-A17B — all under Apache 2.0.
This review covers Qwen 3 models as released on April 28, 2025. Benchmarks cited are from the Qwen Team’s technical report and third-party evaluations published at release. ChatForest reviews AI tools using publicly available information including technical papers, documentation, and community benchmarks. We do not test models hands-on — see our About page for methodology.
Part of our AI Providers and LLM Reviews coverage.