--- title: "Qwen3 ships hybrid thinking and eight open-weight models at once" date: 2026-10-27 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-qwen3-hybrid-reasoning-open-weights language: en --- # Qwen3 ships hybrid thinking and eight open-weight models at once > Alibaba's Qwen3 family pairs thinking and non-thinking modes across two MoE and six dense models, all under Apache 2.0. On April 29, 2025, Alibaba (阿里巴巴) did not release one model — it released a ladder. The Qwen3 (通义千问) family arrived as eight open-weight models spanning 0.6 billion to 235 billion parameters, and the headline feature was not raw size but a switch: each model can think hard or answer fast, on demand. For builders, that single design choice changes how much they pay per token. Qwen3 is Alibaba's first "hybrid reasoning" release. Instead of shipping a separate reasoning model and a separate chat model, one set of weights handles both modes. ## Key takeaways - Qwen3 is eight open-weight (开源权重) models: two MoE (Mixture-of-Experts) and six dense, all released under the Apache 2.0 license. - The flagship Qwen3-235B-A22B has 235 billion total parameters with 22 billion active; the small MoE Qwen3-30B-A3B has 30 billion total with 3 billion active. These are large models (大模型) spanning edge to datacenter. - Every model supports a hybrid "thinking / non-thinking" mode, with controllable reasoning budget up to 38K tokens via the API. - The family covers 119 languages and dialects and was trained on 36 trillion tokens — double the Qwen2.5 dataset. ## A family, not a single model The eight models are: - Dense: Qwen3-0.6B, 1.7B, 4B, 8B, 14B, 32B. - MoE: Qwen3-30B-A3B (30B total / 3B active), Qwen3-235B-A22B (235B total / 22B active). Small dense models run on a phone or a single GPU; the big MoE models serve production traffic. The largest context length is 128K tokens on the 8B-and-up and MoE variants; the smallest run at 32K. All are available on Hugging Face, ModelScope, and Kaggle, and deploy through SGLang, vLLM, Ollama, llama.cpp, and similar tooling. ## Two modes in one model The hybrid thinking design is the part worth understanding. In thinking mode, the model reasons step by step before answering — useful for math, coding, and logic. In non-thinking mode, it answers almost immediately for simpler prompts. A developer calling the API can set a thinking budget, capped at roughly 38,000 tokens, trading depth for latency and cost per request. This matters because reasoning models are expensive: every deliberation token is a token you pay for and wait on. By letting the caller dial reasoning up or down per task, Qwen3 pushes cost control into the application layer. Alibaba reports that even Qwen3-4B rivals the older Qwen2.5-72B-Instruct, and that the small MoE Qwen3-30B-A3B beats the previous QwQ-32B while activating one-tenth the parameters. ## Multilingual and agentic Qwen3 supports 119 languages and dialects, roughly doubling Qwen2.5's reach and aimed squarely at international deployments. On the agentic side, the models natively support the Model Context Protocol (MCP) and strong function calling (Function Calling), which Alibaba describes as leading among open models for complex agent tasks. The training corpus is 36 trillion tokens, twice the size of Qwen2.5's, and Alibaba says the gains show up in reasoning, instruction-following, tool use, and multilingual tasks. The flagship was positioned as competitive with DeepSeek-R1, OpenAI's o1, Grok-3, and Gemini-2.5-Pro on coding, math, and general benchmarks. ## Benchmarks worth noting Alibaba highlighted several public evaluations where Qwen3 leads or competes: - AIME25 for math reasoning. - LiveCodeBench for coding. - BFCL for tool and function calling. - Arena-Hard for instruction-following models. On the flagship, the reported results place it among the top open models and within range of closed front-runners on the hardest sets. As with any vendor release, these are self-reported under Alibaba's own harness, so read them as directional. ## Why this matters for the open ecosystem Releasing eight models at once, all under Apache 2.0, does two things. First, it gives developers a drop-in range from edge to datacenter without leaving one license or one codebase. Second, it keeps the open-weight side of the China model boom competitive with closed labs on both capability and cost. Qwen has become one of the most downloaded model families in the world, and Qwen3 extends that base into hybrid reasoning. The timing mattered. Qwen3 landed weeks after a wave of open reasoning releases, and its Apache 2.0 license — more permissive than the Modified MIT some peers use — made it the default choice for commercial products that needed to ship without per-seat fees. Alibaba also bundled Qwen3 into its own cloud and the ModelScope community, so a developer can move from a Hugging Face download to a managed endpoint without rewriting code. For edge devices, the 0.6B and 1.7B variants fit on phones and embedded hardware, which is why Qwen shows up in offline translators, on-device assistants, and cheap inference boxes across Asia and the Global South. ## Honest limitations - The 38K thinking budget helps, but Qwen3 is not a dedicated long-chain reasoning model; on the hardest o1-class math it trades some peak accuracy for speed and cost. - Benchmark numbers are Alibaba's own; independent third-party harnesses sometimes report different orders. - The smallest dense models (0.6B–4B) are capable for their size but not substitutes for the 235B flagship on hard tasks. - Multilingual coverage at 119 languages is broad, but quality is uneven across lower-resource languages. ## What readers can do now - Pick the smallest Qwen3 that meets your task — start at 4B or 8B locally, move to 235B-A22B only when you hit a wall — to cut inference cost. - Turn on thinking mode only for hard prompts and cap the budget; leave it off for chat and classification to save tokens. - If you build agents, test the native MCP and function-calling support against your tools before paying for a closed model. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-qwen3-hybrid-reasoning-open-weights Free to quote with attribution and a link to the original.