Qwen2.5 32B
A 33 billion-parameterHow many numbers the model learned during training — the usual rough measure of its size. dense model from Qwen Team (Alibaba Cloud). Dense models from 0.5B to 72B trained on 18T tokens.
- Size
- 32.8Bdense
- Memory to run
- ~21 GB estimatesmallest, 8K ctx
- Context
- 128Ktokens
- License
- PermissiveApache License 2.0; MIT License
- Updated
- 13 Sept 2026
Where it runs
Best-fitting version on each reference system, at an 8K contextHow much text the model can consider at once, counted in tokens — roughly ¾ of a word each..
Runs well 11
- A100 80GB server48 tok/s measured
- Dual RTX 3090~24 tok/s estimate
- GB10 mini workstation~3 tok/s estimate
- MacBook Pro M3 Max 64 GB~11 tok/s estimate
- MacBook Pro M4 Max 128 GB~15 tok/s estimate
- Mac mini M4 Pro 48 GB~7 tok/s estimate
- Mac Studio M2 Ultra 192 GB~22 tok/s estimate
- RTX 4090 workstation~32 tok/s estimate
- RTX 5090 workstation61 tok/s measured
- RX 7900 XTX desktop~30 tok/s estimate
- Ryzen AI Max+ 395 mini PC~3 tok/s estimate
Slowly (CPU or offload) 2
- CPU-only Ryzen 9 7950X~2 tok/s estimate
- RTX 4060 Ti 16GB budget build~5 tok/s estimate
Measured throughput
| Quantization | System | Runtime | Context | Measurements | Source |
|---|---|---|---|---|---|
| Qwen2.5 32B Instruct AWQ 4-bit | A100 80GB server (512 GB DDR4) | vLLMcuda · 0.7.3 | 8K | 48 tok/s · Generation throughput 72 GB · Peak memory 5,200 tok/s · Prompt throughput 95 ms · Time to first token | Mutinai illustrative fixtures |
| Qwen2.5 32B Instruct Q4_K_M | RTX 5090 workstation (96 GB DDR5) | llama.cppcuda · b4600 | 4K | 2,900 tok/s · Prompt processing (512) 61 tok/s · Generation (128) | Mutinai illustrative fixtures |
Versions & downloads
Each version is a separate set of weights. Expand one for its downloads: quantizationStoring each of the model’s numbers with fewer bits, so the file is smaller and needs less memory. trades a little quality for a much smaller file, and the memory column adds the working memoryExtra space the model needs while it answers. It grows with the length of the conversation, on top of the file itself. a conversation needs on top.
Variant Qwen2.5 32BPermissive
- Good for
- Publisher
- Qwen Team (Alibaba Cloud)
- License
- Apache License 2.0 · commercial use allowed
- Released
- 19 Sept 2024
- huggingface
- Qwen/Qwen2.5-32B
Contributions are not open yet, so there is nothing here from members.
| Quantization | Format | Bits | Download | Memory @ 8K | Publisher |
|---|---|---|---|---|---|
| Download BF16native | safetensors | 16 | 61.0 GiB | Qwen Team (Alibaba Cloud)Qwen/Qwen2.5-32B |
Variant Qwen2.5 32B InstructPermissive
- Good for
- Publisher
- Qwen Team (Alibaba Cloud)
- License
- Apache License 2.0 · commercial use allowed
- Released
- 19 Sept 2024
- huggingface
- Qwen/Qwen2.5-32B-Instruct
Contributions are not open yet, so there is nothing here from members.
| Quantization | Format | Bits | Download | Memory @ 8K | Publisher |
|---|---|---|---|---|---|
| Download BF16native | safetensors | 16 | 61.0 GiB | Qwen Team (Alibaba Cloud)Qwen/Qwen2.5-32B-Instruct | |
| Download Q5_K_Mk quant | gguf | 5.69 | 21.7 GiB | LM Studio Communitylmstudio-community/Qwen2.5-32B-Instruct-GGUF | |
| Download Q4_K_Mk quant | gguf | 4.89 | 18.7 GiB | Qwen Team (Alibaba Cloud)Qwen/Qwen2.5-32B-Instruct-GGUF | |
| Download AWQ 4-bitawq | safetensors | 4.6 | 17.5 GiB | Qwen Team (Alibaba Cloud)Qwen/Qwen2.5-32B-Instruct-AWQ | |
| Download MLX 4-bitmlx | mlx | 4.5 | 17.2 GiB | MLX Communitymlx-community/Qwen2.5-32B-Instruct-4bit |
Variant DeepSeek-R1-Distill-Qwen-32BPermissive
- Good for
- Publisher
- DeepSeek
- License
- MIT License · commercial use allowed
- Released
- 20 Jan 2025
- huggingface
- deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
Contributions are not open yet, so there is nothing here from members.
| Quantization | Format | Bits | Download | Memory @ 8K | Publisher |
|---|---|---|---|---|---|
| Download BF16native | safetensors | 16 | 61.0 GiB | DeepSeekdeepseek-ai/DeepSeek-R1-Distill-Qwen-32B | |
| Download Q4_K_Mk quant | gguf | 4.89 | 18.7 GiB | LM Studio Communitylmstudio-community/DeepSeek-R1-Distill-Qwen-32B-GGUF | |
| Download MLX 4-bitmlx | mlx | 4.5 | 17.2 GiB | MLX Communitymlx-community/DeepSeek-R1-Distill-Qwen-32B-4bit |
Variant RaDaR 32BPermissive
- Good for
- Publisher
- sczzz
- License
- Apache License 2.0 · commercial use allowed
- Released
- 19 May 2026
- huggingface
- sczzz/RaDaR-32B
Contributions are not open yet, so there is nothing here from members.
No downloadable artifacts recorded.
Other sizes in Qwen2.5: Qwen2.5 7B, Qwen2.5 14B
Benchmarks
Each benchmarkA fixed set of questions every model is given, so their scores can be compared on the same task. score is labelled with who produced it. Developer-reported scores use each lab’s own prompts and settings; a benchmark that runs every model itself uses one setup for all. Bars are relative to the best open result.
Community
Benchmark runs and reviews from people running it on their own hardware.
Benchmark runs
Reviews
Details & sources
Architecture
- Parameters
- 32.76B
- Active / token
- All (dense)
- Architecture
- Dense
- Layers
- 64
- Attention heads
- 40
- KV heads
- 8
- Head dim
- 128
- KV cache @ 8K (fp16)
- 2.00 GiB
- Max context
- 131,072 tokens
Lineage
How this model’s versions relate to each other and to other models.
Sources
- Mutinai illustrative fixtures (illustrative fixture, 1 record, 13 Sept 2026)
- Hugging Face Hub (live source, 1 record, 13 Sept 2026)
External identifiers
- huggingface Qwen/Qwen2.5-32B
- huggingface Qwen/Qwen2.5-32B-Instruct
- huggingface deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
- huggingface sczzz/RaDaR-32B
Field history
- base = "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B" — Hugging Face Hub (current)
- derivation = "fine_tune" — Hugging Face Hub (current)
- license = "apache-2.0" — Hugging Face Hub (current)
- name = "RaDaR 32B" — Hugging Face Hub (current)
Timeline
- Qwen2.5 family released