[whatmodel.sh](../../README.md) / Qwen3 30B-A3B Instruct 2507
# Qwen3 30B-A3B Instruct 2507
Alibaba · released 2025-07 · 30.5B params (3.3B active) · apache-2.0
**score 60.1** ±0.6 · #199 of 732 · from 60 benchmark results across 10 sources
Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only.
## What it's good at
● marks Qwen3 30B-A3B Instruct 2507 among every model on that task or test, worst to best.
task vs the others rank counts tests
agent tasks ▁▁▃▆●▄▅▅▄▄▂▁ #212 of 313 19% 7
coding ▁▁▁▃▅▇█●▅▄▃▁ #209 of 469 16% 3
chat & writing ▁▁▁▃▆▄▆█●▇▇▃ #133 of 402 8.7% 3
math ▁▁▃▅█▆▄●▄▂▁▁ #135 of 431 12% 4
agent tasks · #212 of 313 · 19% of its score
Beats [o1 (2024-12-17)](o1-2024-12-17.md), [o3-mini](o3-mini.md), [DeepSeek R1 (Jan '25)](deepseek-r1.md) here, though they rank higher overall. Trails [SmolLM3 3B](smollm3-3b.md), [Jamba 1.5 Mini](jamba-1-5-mini.md), [Mellum2 12B-A2.5B Instruct](mellum2-12b-a2-5b-instruct.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
τ²-bench Telecom (maker) 12.3% ▁·●█▆▅▂▃▃▂▃▁ beats 2/89 very hard · avg 8.0% 4.8% sharp · at its level · several runs
τ-bench Retail (maker) 59.1% ▁····▃·▂●█▂▆ beats 8/28 hard · avg 32% 4.0% at its level · several runs
τ²-bench Retail (maker) 57.0% ▁····▂▁●▆█▃▁ beats 8/44 medium · avg 35% 2.7% at its level · several runs
τ²-bench Airline (maker) 38.0% ▁▁·●▁▃▂▁█▅▂▃ beats 4/42 hard · avg 22% 2.4% at its level · several runs
τ-bench Airline (maker) 40.0% ▁▁▁·▅▄●█▇▁▁▇ beats 10/28 hard · avg 19% 2.4% at its level · several runs
BFCL V4 (December 2025 archive) 41.4% ▁▁▁▂██●▅▆▂▂▂ beats 47/77 hard · avg 31% 1.8% at its level
BFCL v3 (maker) 65.1% ▁▁▁▁▁▃▅█▆●▄▁ beats 126/171 medium · avg 56% 1.1% blunt · at its level · several runs
coding · #209 of 469 · 16% of its score
Trails [QwQ-32B-Preview](qwq-32b-preview.md), [Ministral 3 3B Reasoning 2512](ministral-3-3b-reasoning.md), [DeepSeek-Coder-V2 Instruct (0724)](deepseek-coder-v2-0724.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
Aider polyglot (225 exercises, tw… 35.6% ▂▁▁▃▆▃●▆█▃▄▂ beats 30/71 very hard · avg 8.2% 11% sharp · at its level · several runs
LiveCodeBench v6 (Feb–May 2025) (… 43.2% ▁▁●▁█▅▄▇▅▅▃█ beats 2/38 hard · avg 27% 2.8% at its level · several runs
MultiPL-E (maker) 83.8% ▃··▄·▅▅▇▄█●▄ beats 28/31 easy · avg 68% 2.4% several runs
chat & writing · #133 of 402 · 8.7% of its score
Beats [LongCat 2.0 (1.6T-A48B)](longcat-2-0.md), [o3-mini](o3-mini.md), [K-EXAONE (236B-A23B)](k-exaone-236b-a23b.md) here, though they rank higher overall. Trails [LFM2.5 8B-A1B](lfm2-5-8b-a1b.md), [LFM2.5 2.6B](lfm2-5-2-6b.md), [Nemotron-Cascade-8B](nemotron-cascade-8b.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
LMArena (text) 1382.3 ▁▁▁▂▃▃▃▆●▇█▃ beats 222/385 avg 1260 3.7% blunt
Arena-Hard v2 aggregate (maker) 69.0% ▁▁··▄▃▁▃█●▁▇ beats 16/22 hard · avg 26% 3.1% sharp · at its level · small run · seve…
IFEval (aggregation unspecified) … 84.7% ▁▂▄▂▂▄▆●█▅▂▁ beats 71/145 easy · avg 76% 1.9% several runs
math · #135 of 431 · 12% of its score
Beats [Claude Sonnet 4.6](claude-sonnet-4-6.md), [MiniMax-M3](minimax-m3.md), [MiMo-V2.5 (310B-A15B)](mimo-v2-5.md) here, though they rank higher overall. Trails [Olmo 3.1 32B Think](olmo-3-1-32b-think.md), [Qwen3-VL 30B-A3B Thinking](qwen3-vl-30b-a3b-thinking.md), [Granite 4.2 8B](granite-4-2-8b.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
HMMT 2025 (edition unspecified) (… 43.0% ▃●▇▅▇▅▁·▁··▁ beats 6/32 very hard · avg 15% 3.5% sharp · at its level · several runs
AIME 2025 (maker) 61.3% ▁·▁▁▁▂▅●█▇▅▁ beats 50/133 hard · avg 25% 3.3% sharp · at its level · several runs
OTIS mock AIME 2024–2025 62.2% ▃▃▄▃▃▃●▇█▃▂▅ beats 90/197 very hard · avg 7.2% 3.0% sharp · at its level · noisy
PolyMATH (maker) 43.1% ▁···▃▁·●█▄▄▄ beats 5/23 hard · avg 24% 2.0% at its level · several runs
other tests (19)
test result vs the others in its run difficulty counts why
GPQA (subset unspecified) (maker) 70.4% ▂▁▂▂▄▃▆▃▅●▃▃ beats 79/106 hard · avg 44% 6.5% sharp · at its level · several runs
GPQA Diamond 55.6% ▁▁▁▁▂▄▆●▅▇█▅ beats 78/211 hard · avg 39% 6.3% sharp · at its level
MMLU-Pro (maker) 78.4% ▁▁·▁▁▁▃▂▄●█▁ beats 118/188 medium · avg 58% 4.1% at its level · several runs
INCLUDE (maker) 71.9% ▂····▂▂▄▄●█▄ beats 18/36 medium · avg 57% 3.5% at its level · several runs
MMLU-Redux (maker) 89.3% ▁▁▃▁▂▃▄▅●▅█▄ beats 33/63 easy · avg 78% 3.4% several runs
Medmarks-V weighted mean win rate… 53.2% ▂▃▅▂▆▃█●█▅▃▆ beats 20/34 medium · avg 41% 3.0% blunt · at its level
BRIDGE clinical text, zero-shot a… 36.9% ▁▁▁▁▁▂▂▃▄●█▃ beats 53/97 hard · avg 31% 2.3% blunt · at its level
Creative Writing v3 (maker) 86.0% ▂····▂▆··▆●▆ beats 10/14 easy · avg 69% 2.3% several runs
MMLU-ProX (maker) 72.0% ▁·▁▁▁▂▁●▃█▅▄ beats 9/36 medium · avg 57% 2.3% at its level · several runs
SuperGPQA (maker) 53.4% ▂▁▂▂▅▆▃●▇▇█▃ beats 24/46 medium · avg 36% 2.2% at its level · several runs
LiveBench 2024-11-25 (maker) 69.0% ▄··●·▂·▂█▆▂▂ beats 2/13 medium · avg 55% 1.8% at its level · several runs
WritingBench (maker) 85.5% ▁·▁▂▁··▅█●▂▁ beats 17/20 easy · avg 70% 1.8% several runs
Multi-IF (aggregation unspecified… 67.9% ▁▁▁▁▁▁●▄█▄▁▂ beats 10/41 medium · avg 59% 1.6% at its level · several runs
Chess Puzzles (Epoch) 2.0% ●▁▁▁▁▂▃▃▃▂▁▁ beats 43/137 very hard · avg 5.8% 0.7% sharp · too hard for it · noisy
Open PT-LLM Leaderboard average (… 78.9% ▃▁▁▁▁▁▁▁▂█●▁ beats 248/276 medium · avg 72% 0.6% noisy
EuroEval HellaSwag (English), tes… 88.3% ▃·▁▁▂▁▂▅█▇●▁ beats 89/100 medium · avg 67% 0.5% sharp · noisy
ZebraLogic (maker) 90.0% █·▁▁▁▇·▄●▁▄▁ beats 17/24 medium · avg 43% 0.5% at its level · noisy · small run
EuroEval Life in the UK (English … 90.6% ▁▁▁▁▁▁▁▂▄█●▃ beats 118/146 easy · avg 82% 0.2% noisy
LIBRA Mini (Russian long context,… 51.5% ▄···▄█·██··● beats 8/8 medium · avg 38% 0.2% blunt · at its level
source results share of the score
Qwen/Qwen3-Next-80B-A3B-Instruct model … 24 43% ▇▇▇▇▇
Qwen/Qwen3-30B-A3B-Instruct-2507 model … 25 35% ▇▇▇▇
Epoch AI 3 10% ▇
LMArena leaderboard dataset 1 4% ▇
Medmarks: A Comprehensive Open-Source L… 1 3% ▇
BRIDGE Medical Leaderboard 1 2% ▇
Berkeley Function Calling Leaderboard r… 1 2% ▇
EuroEval leaderboards 2 1% ▇
Open Portuguese LLM Leaderboard 1 1% ▇
ai-forever LIBRA leaderboard 1 0% ▇
prior: pull toward the middle 0% ▇
weighted average = 60.1
Licences: Qwen/Qwen3-Next-80B-A3B-Instruct model card ([apache-2.0](../../licenses/Qwen--Qwen3-Next-80B-A3B-Instruct.txt)) · Qwen/Qwen3-30B-A3B-Instruct-2507 model card ([apache-2.0](../../licenses/Qwen--Qwen3-30B-A3B-Instruct-2507.txt)) · Epoch AI ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · LMArena leaderboard dataset ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · Medmarks: A Comprehensive Open-Source LLM Benchmark Suite for Medical Tasks ([CC BY 4.0](../../licenses/paper-2605.01417v1-CC-BY-4.0.txt)) · BRIDGE Medical Leaderboard ([MIT](../../licenses/bridge-medical-leaderboard-MIT.txt)) · Berkeley Function Calling Leaderboard result archive ([Apache-2.0](../../licenses/bfcl-result-archive-v4-Apache-2.0.txt)) · EuroEval leaderboards ([MIT](https://github.com/EuroEval/leaderboards/blob/1a77f1d2f8ba1ba05ab73287fb090d76bf41b2e0/LICENSE)) · Open Portuguese LLM Leaderboard ([Apache-2.0](../../licenses/hf-leaderboard-spaces-Apache-2.0.txt)) · ai-forever LIBRA leaderboard ([MIT](../../licenses/hf-leaderboard-spaces-MIT.txt))
## How to run it
### Through an API
api (Nebius) · $0.15/M blended · $0.1 in / $0.3 out per 1M tokens · [tokenfactory.nebius.com](https://tokenfactory.nebius.com/models/catalog/text2text/Qwen%2FQwen3-30B-A3B-Instruct-2507)
[price source](https://github.com/BerriAI/litellm/blob/c19ce71bcdad234ca87a7af10cee45375ce811a8/model_prices_and_context_window.json): USD per 1M text tokens. Nebius standard synchronous API; model Qwen/Qwen3-30B-A3B-Instruct-2507. Uncached input; output includes generated reasoning where billed. Nebius FP8 serving variant; FI / eu-north1. This serving precision is not a claim about benchmark precision. Batch, cache storage, tools and other modalities excluded. · [context source](https://github.com/BerriAI/litellm/blob/c19ce71bcdad234ca87a7af10cee45375ce811a8/model_prices_and_context_window.json): Nebius API maximum input tokens from exact catalogue record nebius/Qwen/Qwen3-30B-A3B-Instruct-2507; not the legacy max\_tokens field.
### llama.cpp (recommended)
1. Install llama.cpp. Homebrew and winget installs name the commands llama-cli and llama-server instead. [Other ways to install](https://github.com/ggml-org/llama.cpp/blob/master/docs/install.md)
```sh
# macOS / Linux
curl -LsSf https://llama.app/install.sh | sh
# Windows
irm https://llama.app/install.ps1 | iex
```
2. Run Qwen3 30B-A3B Instruct 2507 (Q4_K_M, needs ~18.6 GB, from [huggingface.co/lmstudio-community](https://huggingface.co/lmstudio-community/Qwen3-30B-A3B-Instruct-2507-GGUF/tree/02db009a02d5f8817f8a79e952cb0e8d79e9bd34))
```sh
llama cli -hf lmstudio-community/Qwen3-30B-A3B-Instruct-2507-GGUF:Q4_K_M
```
3. Or start a local server and open the address it prints (http://localhost:9931)
```sh
llama serve -hf lmstudio-community/Qwen3-30B-A3B-Instruct-2507-GGUF:Q4_K_M
```
Other versions: Q8_0 (~32.5 GB) `llama cli -hf lmstudio-community/Qwen3-30B-A3B-Instruct-2507-GGUF:Q8_0`
### LM Studio
1. Install LM Studio. Open the app once before using the lms command. [Download it (macOS / Windows / Linux)](https://lmstudio.ai/download)
2. Run Qwen3 30B-A3B Instruct 2507 (Q4_K_M, needs ~18.6 GB, from [huggingface.co/lmstudio-community](https://huggingface.co/lmstudio-community/Qwen3-30B-A3B-Instruct-2507-GGUF/tree/02db009a02d5f8817f8a79e952cb0e8d79e9bd34))
```sh
lms get https://huggingface.co/lmstudio-community/Qwen3-30B-A3B-Instruct-2507-GGUF@Q4_K_M -y
```
3. Then run lms ls to see its name, and lms chat followed by that name
Other versions: Q8_0 (~32.5 GB) `lms get https://huggingface.co/lmstudio-community/Qwen3-30B-A3B-Instruct-2507-GGUF@Q8_0 -y` · 4bit (~17.2 GB, Mac only) `lms get https://huggingface.co/lmstudio-community/Qwen3-30B-A3B-Instruct-2507-MLX-4bit -y` · 8bit (~32.4 GB, Mac only) `lms get https://huggingface.co/lmstudio-community/Qwen3-30B-A3B-Instruct-2507-MLX-8bit -y`
### Ollama
1. Install Ollama. [Download it (macOS / Windows)](https://ollama.com/download)
```sh
# Linux
curl -fsSL https://ollama.com/install.sh | sh
```
2. Run Qwen3 30B-A3B Instruct 2507 (Q4_K_M, needs ~18.8 GB, from [ollama.com](https://ollama.com/library/qwen3/tags))
```sh
ollama run qwen3:30b-instruct
```
Other versions: Q8_0 (~37.1 GB) `ollama run qwen3:30b-a3b-instruct-2507-q8_0`
### MLX (Mac)
1. Install MLX (Mac). Apple silicon Macs only, text only.
```sh
# with uv
uv tool install mlx-lm
# or
pip install mlx-lm
```
2. Run Qwen3 30B-A3B Instruct 2507 (4bit, needs ~17.2 GB, Mac only, from [huggingface.co/lmstudio-community](https://huggingface.co/lmstudio-community/Qwen3-30B-A3B-Instruct-2507-MLX-4bit/tree/22d11bad31cf8f3e01acd9fbb086026e143e5778))
```sh
mlx_lm.chat --model lmstudio-community/Qwen3-30B-A3B-Instruct-2507-MLX-4bit
```
3. Or start an OpenAI-compatible server at http://localhost:8080/v1
```sh
mlx_lm.server --model lmstudio-community/Qwen3-30B-A3B-Instruct-2507-MLX-4bit
```
Other versions: 8bit (~32.4 GB, Mac only) `mlx_lm.chat --model lmstudio-community/Qwen3-30B-A3B-Instruct-2507-MLX-8bit`
Model repository and documentation: [huggingface.co/Qwen](https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507)
## Published speed measurements
Reference results for the hardware, engine and settings below. These are not verified speeds of the downloads above or predictions for your machine.
hardware precision speed engine and settings
Apple M4 Max (64GB; CPU/GPU… MLX 8-bit affin… 83.16 tok/s mlx-lm 0.28.2, mlx 0.29.2.dev2025…
Apple M4 Max (64GB; CPU/GPU… MLX 6-bit affin… 94.14 tok/s mlx-lm 0.28.2, mlx 0.29.2.dev2025…
Apple M4 Max (64GB; CPU/GPU… MLX 5-bit affin… 101 tok/s mlx-lm 0.28.2, mlx 0.29.2.dev2025…
Apple M4 Max (64GB; CPU/GPU… MLX 4-bit affin… 113.33 tok/s mlx-lm 0.28.2, mlx 0.29.2.dev2025…
AMD Ryzen AI Max+ 395 (Stri… IQ4_XS-3.63bpw 100.38 tok/s llama.cpp ac4cddeb0 (build_number…
AMD Ryzen AI Max+ 395 (Stri… IQ4_XS-3.63bpw 98.32 tok/s llama.cpp 2016bf2 (build_number 1…
AMD Ryzen AI Max+ 395 (Stri… IQ4_XS-3.63bpw 100.04 tok/s llama.cpp 1fd5f4803 (build_number…
AMD Ryzen AI Max+ 395 (Stri… IQ4_XS-3.63bpw 103.18 tok/s llama.cpp 98d5e8ba8 (build_number…
Each speed links to its source, which has the full settings.
## Quantization evidence
How much of the full-precision result a quantized version kept, measured within one study. Rows from the same study share a table and aren't independent. They don't describe every download of that precision.
precision test full → quantized kept
GGUF Q2_K_S produced by AutoR… Average accuracy of 10 lm-eval ta… 0.675 → 0.639 94.7%
GGUF Q2_K_S produced by AutoR… Average accuracy of 10 lm-eval ta… 0.675 → 0.646 95.8%
GGUF Q3_K_S produced by AutoR… Average accuracy of 10 lm-eval ta… 0.675 → 0.66 97.8%
GGUF Q3_K_S produced by AutoR… Average accuracy of 10 lm-eval ta… 0.675 → 0.662 98.2%
GGUF Q4_K_S produced by AutoR… Average accuracy of 10 lm-eval ta… 0.675 → 0.676 100.2%
GGUF Q4_K_S produced by AutoR… Average accuracy of 10 lm-eval ta… 0.675 → 0.678 100.5%
INT4 W4A16 (weight-only INT4,… MATH-500 — published row 'Math 50… 0.899 → 0.922 102.5%
INT4 W4A16 (weight-only INT4,… IFEval — published row 'IfEval'; … 0.891 → 0.891 100.0%
INT4 W4A16 (weight-only INT4,… GSM8K-Platinum — published row 'G… 0.961 → 0.958 99.7%
INT4 W4A16 (weight-only INT4,… MMLU (CoT) — published row 'MMLU-… 0.843 → 0.835 99.1%
INT4 W4A16 (weight-only INT4,… MMLU-Pro — published row 'MMLU-Pr… 0.789 → 0.771 97.7%
INT8 W8A8 (weights INT8 per-c… GSM8K-Platinum — published row 'G… 0.961 → 0.976 101.5%
INT8 W8A8 (weights INT8 per-c… MATH-500 — published row 'Math 50… 0.899 → 0.905 100.6%
INT8 W8A8 (weights INT8 per-c… MMLU (CoT) — published row 'MMLU-… 0.843 → 0.843 100.0%
INT8 W8A8 (weights INT8 per-c… MMLU-Pro — published row 'MMLU-Pr… 0.789 → 0.788 99.9%
INT8 W8A8 (weights INT8 per-c… IFEval — published row 'IfEval'; … 0.891 → 0.889 99.7%
MLX 4-bit affine quantization… MMLU-Pro (mlx_lm.evaluate --task … 0.726 → 0.707 97.4%
MLX 5-bit affine quantization… MMLU-Pro (mlx_lm.evaluate --task … 0.726 → 0.72 99.1%
MLX 6-bit affine quantization… MMLU-Pro (mlx_lm.evaluate --task … 0.726 → 0.724 99.7%
MLX 8-bit affine quantization… MMLU-Pro (mlx_lm.evaluate --task … 0.726 → 0.725 99.8%
Each kept value links to its source.
[interactive version](https://whatmodel.sh/model/qwen3-30b-a3b-instruct-2507) · [JSON](https://whatmodel.sh/api?format=json&all=1)