Decision models · Skywork

Skywork Reward V2 Llama 3.1 8B

Name: Skywork Reward V2 Llama 3.1 8B
Author: Skywork

Task: Judging and grading answers
Returns: Probability or score
Precision tier: 90 / 100
Size: 7.5B
Max input: 128K tokens
License: llama3.1
Released: 2025-07

What it decides

Returns: One reward score per answer; compare scores to pick the better answer.
Checks: Answer quality and preference
Languages: Multilingual
Precision: 90 / 100 · RewardBench 96.4, RewardBench 2 84.1, JudgeBench 80.0 (model card).
Clarity: Probability or score (15 of 15 clarity points)
How it runs: Reward score head · reads the input once and writes nothing, so it answers in milliseconds.
License: llama3.1 (Custom (read the terms))

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

How to run it

Original weights · Hugging Face

Full-precision files from Skywork, for Transformers, Sentence Transformers, Text Embeddings Inference, Infinity or vLLM.

huggingface.co/Skywork/Skywork-Reward-V2-Llama-3.1-8B (opens in a new tab)

No sign-in needed.

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

Apps that run Skywork Reward V2 Llama 3.1 8B

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)

Sentence Transformers

A few lines of Python with Transformers or Sentence Transformers; easiest for scripts and notebooks.

Windows PC, NVIDIA GPU

Sentence Transformers

A few lines of Python with Transformers or Sentence Transformers; easiest for scripts and notebooks.

Also good: vLLM, SGLang

Windows PC, AMD Radeon

Sentence Transformers

A few lines of Python with Transformers or Sentence Transformers; easiest for scripts and notebooks.

Also good: vLLM, SGLang

Linux, NVIDIA GPU

vLLM

Serves the scoring head directly (rerank, classify or reward) with batching.

Also good: Sentence Transformers, SGLang

Linux server, 80 GB+ GPUs

vLLM

Serves the scoring head directly (rerank, classify or reward) with batching.

Also good: SGLang, Sentence Transformers

App	Type	Opens it as	Systems	Engine	Start
vLLM	Server	Original weights	WSLWindowsLinux	vLLM	`vllm serve Skywork/Skywork-Reward-V2-Llama-3.1-8B`
SGLang	Server	Original weights	WSLWindowsLinux	SGLang	`python3 -m sglang.launch_server --model-path Skywork/Skywork-Reward-V2-Llama-3.1-8B`
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 precision 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
`Q8_0`	10.2 GB	144 ms	6.9/s	90Excellent quality (S) · 90 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 92/100 (quality, fit, speed and memory use) sets the ranking.	All files (opens in a new tab)
`F16`	17.3 GB	165 ms	6.0/s	90Excellent quality (S) · 90 of 100: how capable the model is at F16, whatever it runs on. Match score 92/100 (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 Skywork Reward V2 Llama 3.1 8B

The least expensive GPU in our catalog that runs it without offloading is the Intel Arc B580 (launch price $249) at Q8_0.

Quality	Hardware	Quant	Memory	Time per decision	Runs
90Excellent quality (S) · 90 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 95/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA RTX PRO 6000 Blackwell	`Q8_0`	10.4 GB / 96.0 GB	97 ms	Runs great
90Excellent quality (S) · 90 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 94/100 (quality, fit, speed and memory use) sets the ranking.	AMD Radeon RX 7900 XTX	`Q8_0`	10.4 GB / 24.0 GB	114 ms	Runs great
90Excellent quality (S) · 90 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 94/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA GeForce RTX 5090	`Q8_0`	10.4 GB / 32.0 GB	115 ms	Runs great
90Excellent quality (S) · 90 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 93/100 (quality, fit, speed and memory use) sets the ranking.	AMD Radeon RX 9070 XT	`Q8_0`	10.4 GB / 16.0 GB	142 ms	Runs well
90Excellent quality (S) · 90 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 92/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA GeForce RTX 4090	`Q8_0`	10.4 GB / 24.0 GB	144 ms	Runs well
90Excellent quality (S) · 90 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 90/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA GeForce RTX 5080	`Q8_0`	10.4 GB / 16.0 GB	210 ms	Runs well
90Excellent quality (S) · 90 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 90/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA GeForce RTX 4080 SUPER	`Q8_0`	10.4 GB / 16.0 GB	226 ms	Runs well
90Excellent quality (S) · 90 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 90/100 (quality, fit, speed and memory use) sets the ranking.	AMD Ryzen AI Max+ 395 (128 GB)	`Q8_0`	10.4 GB / 96.0 GB	231 ms	Runs well
90Excellent quality (S) · 90 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 89/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA GeForce RTX 4070 Ti SUPER	`Q8_0`	10.4 GB / 16.0 GB	267 ms	Runs well
90Excellent quality (S) · 90 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 89/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA GeForce RTX 5070 Ti	`Q8_0`	10.4 GB / 16.0 GB	268 ms	Runs well

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.