Core capabilities
- 30B total / about 3B active MoE
- 128K context and turn-level thinking switch
- Multilingual, code, math and tool-use support
- Broad local runtime ecosystem including llama.cpp and MLX
← Model indexMODEL PROFILE / qwen3-30b-a3b
Alibaba Qwen / open
Qwen3 30B-A3B
An efficient MoE model activating roughly 3B parameters while retaining reasoning and agent capabilities.
Context128K
ModalityText
AccessWeights · API
ReasoningHybrid
Where it fits.
Useful when hardware is constrained but hybrid reasoning, tools and multilingual support still matter. Qwen3 30B-A3B activates about 3B of 30B total parameters across a 128-expert MoE, offering 128K context and hybrid thinking in a workstation-friendly tier.
Best-fit workflows
- Workstation deployment
- Lightweight agents
- Multilingual automation
04 / OPERATIONSFrom evaluation
From evaluation
to production.
Deployment notes
Good for local assistants and lightweight agents. Select a runtime with explicit Qwen3 MoE and reasoning-parser support, size memory for all weights plus KV cache, and use higher-tier fallback for hard cases.
Watch before choosing
Local performance depends on the runtime's MoE, KV-cache and reasoning-parser support. Small active parameter count does not mean every 3B runtime can host it; all expert weights still require storage and memory planning. Runtime support for MoE and KV cache is decisive.
05
Family & versions
- Qwen3 235B-A22BFlagship
- Qwen3 30B-A3BEfficient
Specifications and positioning checked against provider documentation.
Open official source↗