Decision models · OpenCompass

CompassJudger-2 32B

Name: CompassJudger-2 32B
Author: OpenCompass

Task: Judging and grading answers
Returns: Written feedback, then a score
Precision tier: 86 / 100
Size: 33B
Max input: 32K tokens
License: apache-2.0
Released: 2025-07

What it decides

Returns: A judgment with reasoning, then a score or a preferred answer.
Checks: Scoring and pairwise comparison
Languages: English and Chinese
Precision: 86 / 100 · Evaluated on JudgerBenchV2 (10 scenarios, 10K questions).
Clarity: Written feedback, then a score (9 of 15 clarity points) · writes about 150 tokens before the answer, so it is slower but explains itself.
How it runs: Writes a short answer · a small language model, so it runs in the same apps as LLMs.
License: apache-2.0 (Open (Apache, MIT))

Source: model card on Hugging Face (opens in a new tab) · How decision models are rated

How to run it

Quantized files · GGUF

Works in LM Studio, llama.cpp, Jan and Ollama. No sign-in needed.

huggingface.co/mradermacher/CompassJudger-2-32B-Instruct-GGUF (opens in a new tab)

ollama run hf.co/mradermacher/CompassJudger-2-32B-Instruct-GGUF

Original weights · Hugging Face

Full-precision files from OpenCompass, for fine-tuning, vLLM, Transformers.

huggingface.co/opencompass/CompassJudger-2-32B-Instruct (opens in a new tab)

No sign-in needed.

Links checked 3 days ago against Hugging Face, Ollama and GitHub.

Apps that run CompassJudger-2 32B

Our pick for each kind of computer, assuming the model fits in its memory. Pick your own GPU or Mac in the Model Finder for an exact answer.

Mac (Apple silicon)

LM Studio

Point-and-click: search, download and chat in one window.

Also good: llama.cpp, Ollama

Windows PC, NVIDIA GPU

LM Studio

Point-and-click: search, download and chat in one window.

Also good: llama.cpp, Ollama

Windows PC, AMD Radeon

Lemonade

AMD’s own app: llama.cpp tuned for Radeon on Vulkan and ROCm.

Also good: LM Studio, llama.cpp

Linux, NVIDIA GPU

LM Studio

Point-and-click: search, download and chat in one window.

Also good: llama.cpp, Ollama

Linux server, 80 GB+ GPUs

vLLM

The full model fits in GPU memory at 8-bit, so it can batch many requests: best for serving a team or an app.

Also good: LM Studio, SGLang

App	Type	Opens it as	Systems	Engine	Start
LM Studio	Desktop app	GGUF	macOSWindowsLinux	llama.cpp, MLX	Open in LM Studio
Ollama	CLI + API	GGUF from Hugging Face	macOSWindowsLinux	llama.cpp, MLX	`ollama run hf.co/mradermacher/CompassJudger-2-32B-Instruct-GGUF:Q4_K_M`
llama.cpp	CLI + API	GGUF	macOSWindowsLinux	llama.cpp	`llama-server -hf mradermacher/CompassJudger-2-32B-Instruct-GGUF:Q4_K_M`
Jan	Desktop app	GGUF	macOSWindowsLinux	llama.cpp, MLX	Open in Jan
vLLM	Server	Original weights	WSLWindowsLinux	vLLM	`vllm serve opencompass/CompassJudger-2-32B-Instruct`
SGLang	Server	Original weights	WSLWindowsLinux	SGLang	`python3 -m sglang.launch_server --model-path opencompass/CompassJudger-2-32B-Instruct`
Docker Model Runner	CLI + API	GGUF	macOSWindowsLinux	llama.cpp, vLLM	`docker model run hf.co/mradermacher/CompassJudger-2-32B-Instruct-GGUF:Q4_K_M`
Lemonade	Desktop app	GGUF	macOSWindowsLinux	llama.cpp, ONNX Runtime GenAI	`lemonade pull mradermacher/CompassJudger-2-32B-Instruct-GGUF:Q4_K_M`
Unsloth Desktop	Desktop app	GGUF	macOSWindowsLinux	llama.cpp, MLX, stable-diffusion.cpp	Open in Unsloth Desktop
Atomic Chat	Desktop app	GGUF	macOSWindowsLinux	llama.cpp, MLX-VLM	Open in Atomic Chat
LocalAI	Server	GGUF	macOSWSLWindowsLinux	llama.cpp, rerankers, Transformers	Get LocalAI
Sentence Transformers	Library	Original weights	macOSWindowsLinux	PyTorch	Get Sentence Transformers

Which apps can open a model depends on the files published for it: GGUF for llama.cpp-based apps, MLX for Apple’s engine, the original weights for vLLM and SGLang. Compare every app.

Memory and speed

Each quantization on an NVIDIA GeForce RTX 4090, for one decision on a 1,024-token input. The best hardware table below covers other GPUs and Macs.

Format	Memory needed	Time per decision	Decisions per second	Quality	Download
`Q4_K_M`	21.1 GB	4.5 s	0.2/s	81Great quality (A) · 81 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 71/100 (quality, fit, speed and memory use) sets the ranking.	GGUF (opens in a new tab)
`Q5_K_M`	24.4 GB	10.1 s	0.1/s	83Great quality (A) · 83 of 100: how capable the model is at Q5_K_M, whatever it runs on. Match score 58/100 (quality, fit, speed and memory use) sets the ranking.	GGUF (opens in a new tab)
`Q6_K`	28.0 GB	22.0 s	0.0/s	85Excellent quality (S) · 85 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 58/100 (quality, fit, speed and memory use) sets the ranking.	GGUF (opens in a new tab)
`Q8_0`	35.9 GB	50.0 s	0.0/s	86Excellent quality (S) · 86 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 52/100 (quality, fit, speed and memory use) sets the ranking.	GGUF (opens in a new tab)
`F16`	66.7 GB	—	—	86Excellent quality (S) · 86 of 100: how capable the model is at F16, whatever it runs on. Match score 11/100 (quality, fit, speed and memory use) sets the ranking.	Original weights (opens in a new tab)

One request at a time. Servers that batch (Text Embeddings Inference, vLLM, Infinity) handle many more decisions per second.

Best hardware for CompassJudger-2 32B

The least expensive GPU in our catalog that runs it without offloading is the Intel Arc Pro B70 (launch price $949) at Q5_K_M.

Quality	Hardware	Quant	Memory	Time per decision	Runs
85Excellent quality (S) · 85 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 77/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA B200	`Q6_K`	37.0 GB / 180 GB	1.0 s	Slow
85Excellent quality (S) · 85 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 76/100 (quality, fit, speed and memory use) sets the ranking.	AMD Instinct MI350X	`Q6_K`	42.4 GB / 288 GB	1.2 s	Slow
83Great quality (A) · 83 of 100: how capable the model is at Q5_K_M, whatever it runs on. Match score 75/100 (quality, fit, speed and memory use) sets the ranking.	AMD Instinct MI325X	`Q5_K_M`	37.2 GB / 256 GB	1.3 s	Slow
81Great quality (A) · 81 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 75/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA H200 141GB SXM	`Q4_K_M`	28.2 GB / 141 GB	1.2 s	Slow
81Great quality (A) · 81 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 75/100 (quality, fit, speed and memory use) sets the ranking.	AMD Instinct MI300X	`Q4_K_M`	30.7 GB / 192 GB	1.2 s	Slow
81Great quality (A) · 81 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 75/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA H200 NVL	`Q4_K_M`	28.2 GB / 141 GB	1.2 s	Slow
86Excellent quality (S) · 86 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 74/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA RTX PRO 6000 Blackwell	`Q8_0`	40.7 GB / 96.0 GB	4.8 s	Very slow
83Great quality (A) · 83 of 100: how capable the model is at Q5_K_M, whatever it runs on. Match score 74/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA GeForce RTX 5090	`Q5_K_M`	26.0 GB / 32.0 GB	3.2 s	Very slow
86Excellent quality (S) · 86 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 74/100 (quality, fit, speed and memory use) sets the ranking.	Apple M3 Ultra (256 GB)	`Q8_0`	51.2 GB / 192 GB	13.1 s	Very slow
86Excellent quality (S) · 86 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 74/100 (quality, fit, speed and memory use) sets the ranking.	Apple M3 Ultra (512 GB)	`Q8_0`	66.6 GB / 384 GB	13.1 s	Very slow

Check it on your own hardware

Estimates come from published hardware specs and our open formulas, not from running the model on your machine. Real speed varies with runtime, drivers and settings. Numbers marked “Measured” come from a linked benchmark.