LOCAL_AI_STACK
v2.2 · 2026-06-12
← back to models

Qwen3 30B-A3B

qwen3-30b-a3b
REASONING

Qwen3 30B-A3B is a 30.5B reasoning model profile for local AI planning. This page records provider, license, context, quantization, hardware-fit estimates, setup hints, source links, and caveats so readers do not choose a model by name alone.

PROVIDER
Qwen
FAMILY
Qwen3
MODEL TYPE
reasoning
PARAMETERS
30.5B (3.3B active)
MODALITIES
text
ARCHITECTURE
Qwen3 mixture-of-experts decoder-only causal LM with 128 experts, 8 activated experts, grouped-query attention, and hybrid thinking/non-thinking chat template.
CONTEXT WINDOW
32,768 tokens
TRAINING TOKENS
36.0T
RELEASE DATE
2025-04-29
LICENSE & LINKS
GGUF REPOSITORIES
Official Qwen GGUF quantization repository for the MoE model; file size reflects total stored parameters, not just active parameters.
SETUP HINTS
OLLAMA
Use `ollama run qwen3:30b`; Ollama's page labels the current packaged 30B as the new 30B model and discusses Qwen3-30B-A3B in the readme.
LM STUDIO
Search for `Qwen/Qwen3-30B-A3B-GGUF` in LM Studio and select Q4_K_M or Q5_K_M.
LLAMA.CPP
Use a recent llama.cpp build with Qwen3-MoE support, for example `llama-server -hf Qwen/Qwen3-30B-A3B-GGUF:Q4_K_M`; control thinking mode with `/think` and `/no_think`.
QUANTIZATION FILE SIZES (GGUF)
Q4_K_M
18.6 GB
Q5_K_M
21.7 GB
Q8_0
32.5 GB
Quantization note: Q4_K_M for practical local use; Q5_K_M if 32GB+ unified memory or ample VRAM is available.
RAM / VRAM ESTIMATES
MIN RAM
23.2 GB
COMFORTABLE RAM
29.3 GB
MIN VRAM
22.3 GB
COMFORTABLE VRAM
28.2 GB
These are conservative local-inference estimates, not official hardware requirements.
HARDWARE FIT
CPU only (no GPU)
possible_but_slow
8 GB RAM
not_recommended
16 GB RAM
not_recommended
32 GB RAM
good
8 GB VRAM
not_recommended
12 GB VRAM
not_recommended
24 GB VRAM
usable
Apple Silicon (unified memory)
good_32gb_plus
Hardware fit values are conservative local-inference estimates based on GGUF size plus runtime and KV-cache overhead. Actual requirements depend on context length, quantization, runtime, GPU offload, and other running apps.
BEST FOR
·Strong local reasoning when you want more capability than dense 14B without running a full 32B dense model at every token.
·Math, coding, and logical reasoning workloads that benefit from Qwen3 thinking mode.
·Local agent experiments where Qwen3 tool/agent behavior is useful and memory can handle the full MoE checkpoint.
AVOID IF
·You need an 8GB or 16GB RAM laptop target.
·You need the smallest possible download; MoE still stores all experts on disk.
·Your runtime lacks stable Qwen3-MoE support.
CAVEATS
·RAM and VRAM figures are estimates from GGUF file size plus conservative runtime and KV-cache overhead; exact memory depends on context length, batch size, backend, and GPU offload.
·For MoE models, activeParametersB reflects the official activated-parameter count per token, while download and memory needs reflect the full stored parameter set.
·Qwen3 model cards state 32,768 native context and 131,072 tokens with YaRN; this record uses the native context as the primary contextWindowTokens value.
·Ollama currently presents newer Qwen3 30B packaging as `qwen3:30b`; the official Hugging Face artifact for this requested target is `Qwen3-30B-A3B`.
SOURCE URLS
FIELD EVIDENCE
LANGUAGES
ChineseEnglishFrenchSpanishPortugueseGermanItalianRussianJapaneseKoreanVietnameseThaiArabic100+ languages and dialects
CAPABILITIES
tool callingreasoning tuned
← Qwen3 14BDeepSeek-R1-Distill-Qwen-14B →