01Overview02Models03Rankings04Compare05Quiz06Guide

Model

Qwen3-30B-A3B-Instruct-2507 NO.34

Small, fast open-weight MoE with big-model manners.

CompareBack to the index

Alibaba Qwen · Qwen · Fast · Open weights

Specification

Fig.01 — Qwen — Alibaba Qwen

CHINA

Context window
262Ktokens
Max output
33Ktokens
Input price
$0.05/ 1M
Output price
$0.19/ 1M
Cached input
$0.05/ 1M
Throughput
—
Class
Fast
Modalities
Text
Weights
Open
Parameters
30B
License
Apache-2.0
Knowledge cutoff
Apr 2025
Released
Jul 2025

Overview

Qwen3-30B-A3B-Instruct-2507 is a compact open-weight MoE (30B total, 3B active) tuned for fast, efficient instruction following. It is remarkably capable for its size, runs cheaply, and is easy to self-host under Apache-2.0. It is not built for complex long-horizon agentic or coding work, where its limits show quickly.

Strengths

  • Very fast and cheap thanks to 3B active params
  • Great capability per active parameter
  • Runs on consumer-grade GPUs
  • Apache-2.0 open weights

Trade-offs

  • Limited depth on hard reasoning
  • Not suited to complex coding
  • Text-only, no vision
  • Struggles with long agentic chains

Fit

Best for

  • On-device and edge deployment
  • High-volume simple tasks
  • Routing, classification, and lightweight assistants
  • Cost-optimized chat backends

Not ideal for

  • Complex agentic coding
  • Hard multi-step reasoning

Capabilities

Ten axes, normalized 0–100 and scored the same way across the whole catalog — so a 78 here means what a 78 means anywhere else on the bench.

Fig.02 — Capability radar

0 — 100

Reasoning
66
Coding
60
Math
68
Writing
70
Knowledge
72
Speed
90
Agentic
58
Vision
5
Multilingual
82
Long Context
74

Benchmarks

Public results, with independent third-party runs marked. Bars normalize percentages against 100 and Elo ratings against a 1500 ceiling.

BenchmarkResult
MMLU-ProKnowledge · Vendor-reported
78.4%
GPQA DiamondIndependentReasoning · Independent aggregators
70.4%
AIME 2025Math · Vendor-reported
61.3%
LiveCodeBenchIndependentCoding · LiveCodeBench
43.2%
LiveBenchIndependentGeneral · LiveBench
74.3%

Alternatives

Models in roughly the same class — the ones worth weighing against this record.

Compare side by side
01Mistral Small 4Mistral AITiny Apache-2.0 multimodal model unifying reasoning, vision and coding.02Qwen3-Coder-480B-A35BAlibaba QwenBest open-weights Qwen for agentic coding — great value.03GPT-5.4 nanoOpenAIOpenAI's cheapest small model for high-volume simple work.

Index

PreviousQwen3-235B-A22B-Thinking-2507NextQwen3-Coder-480B-A35B
All models/EDUCATION