Curiosity · AI Model

Skywork-o1-Open

Skywork-o1-Open is Kunlun Tech's open-weight reasoning LLM family, released in late 2024. It comes in 8B and 32B sizes and was one of the first large-scale open reproductions of OpenAI's o1 reasoning paradigm, combining rejection-sampling fine-tuning with process reward models to produce explicit chain-of-thought traces.

Model specs

Vendor
Skywork
Family
Skywork-o1
Released
2024-11
Context window
8,192 tokens
Modalities
text

Strengths

  • Strong math scores for an open-weight 32B model
  • Comes in two sizes for different compute budgets
  • Process reward model training recipe is publicly documented

Limitations

  • Smaller max context than frontier models
  • Surpassed by DeepSeek-R1 on hardest competition math
  • English-first training; Chinese reasoning is good but not dominant

Use cases

  • Open-weight o1 reproductions for research
  • Math tutoring apps needing visible chain-of-thought
  • Baselines for building custom reasoning LLMs
  • Teaching process reward modelling

Benchmarks

BenchmarkScoreAs of
MATH~73%2026-04
GSM8K~95%2026-04
AIME 2024~20%2026-04

Frequently asked questions

What is Skywork-o1-Open?

Skywork-o1-Open is Kunlun Tech's open-weight family of o1-style reasoning language models, available in 8B and 32B sizes. It uses process reward models and rejection sampling to teach explicit step-by-step reasoning.

How does Skywork-o1-Open compare to Marco-o1?

Marco-o1 relies on Monte Carlo Tree Search at inference time, while Skywork-o1-Open is trained end-to-end with process rewards. Skywork's 32B variant generally scores higher, at the cost of heavier hardware.

Sources

  1. Skywork-o1-Open on HuggingFace — accessed 2026-04-20
  2. Skywork AI homepage — accessed 2026-04-20