The Open-Source AI Stack
RSS
All models

Models · qwen

Qwen 3 32B Instruct

Open Alibaba · 2025-04-28 · Apache-2.0

Dense 32B variant of the Qwen 3 release, shipped alongside the MoE 235B A22B flagship and a full size ladder from 0.6B to 32B dense. All sizes ship under Apache 2.0, with the same hybrid thinking-vs-fast inference toggle as the MoE.

Cost

$0.15 / Mtok input
$0.59 / Mtok output

Together AI · as of 2026-05-21

via Artificial Analysis ↗

Speed

98.7 tok/sec output
1200 ms TTFT

· as of 2026-05-21

source ↗

Architecture

tokens in Embedding vocab not disclosed · qwen tokenizer × 64 layers Grouped-Query Attention RoPE context 131,072 tokens Dense MLP SwiGLU activation (standard) 32.8B active params Output projection tokens out
Schema-generated from data/models.yaml. Every label is auditable against the model's sources.

Specs

Architecture
dense
Total params
32.8B
Active params
32.8B
Context window
131K tokens
Attention
gqa
Position encoding
rope
Pretraining tokens
36.0T
Post-training
sft, grpo, rlhf
OSI-approved
yes
Data released
no
Training code
not released

Benchmarks

Each score carries the date it was published; we never infer or interpolate missing scores.

General reasoning

MMLU-Pro 72.7 as of 2026-05-21 source ↗
GPQA-Diamond 53.5 as of 2026-05-21 source ↗

Code

LiveCodeBench 28.8 as of 2026-05-21 source ↗

Math

MATH 86.9 as of 2026-05-21 source ↗
AIME 2024 30.3 as of 2026-05-21 source ↗
AIME 2025 19.7 as of 2026-05-21 source ↗

Available quantizations

GGUF llama.cpp's container; the common local format, k-quants from Q2 to Q8. runs on llama.cpp, Ollama
AWQ Activation-aware 4-bit weight quantization for GPU serving. runs on vLLM, SGLang
GPTQ Post-training 4-bit weight quantization for GPU serving. runs on vLLM, SGLang, Transformers
EXL2 ExLlamaV2's variable-bitrate format for consumer GPUs. runs on ExLlamaV2
MLX Apple MLX 4/8-bit layout for Apple silicon. runs on Apple MLX
FP8 8-bit float, frequently a native release on Hopper / Blackwell GPUs. runs on vLLM, SGLang, TensorRT-LLM
bitsandbytes On-the-fly NF4 / INT8 weight quantization inside Transformers. runs on Transformers

Verified via the Hugging Face model tree ↗. Community quantizations change over time; the families shown are those with published weights at audit time.

Notable innovations

  • · Hybrid thinking mode toggle
  • · Four-stage post-training including long-CoT cold start
  • · 131K context via YaRN, 32K native
  • · Apache 2.0 across the full Qwen 3 size ladder

Sources