Skywork/Skywork-o1-Open-Llama-3.1-8B

Hugging Face
TEXT GENERATIONPricing:Input $0.2 / Cached $0.028 / Output $0.32Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Nov 26, 2024License:otherArchitecture:Transformer0.1K Featherless Exclusive Warm

Skywork/Skywork-o1-Open-Llama-3.1-8B is an 8 billion parameter chat model developed by the Skywork team at Kunlun Inc., built upon the Llama-3.1-8B architecture. This model is specifically enhanced with "o1-style" data and a three-stage training scheme to significantly improve reasoning capabilities, including reflective reasoning, process reward modeling, and online reasoning planning. It excels in complex problem-solving across common-sense, logical, mathematical, and coding challenges, demonstrating advanced thinking, planning, and self-reflection abilities.

Loading preview...

Skywork o1 Open-Llama-3.1-8B: Enhanced Reasoning Model

Skywork/Skywork-o1-Open-Llama-3.1-8B is an 8 billion parameter chat model from the Skywork team at Kunlun Inc., designed to integrate "o1-like" slow thinking and reasoning. Built on the Llama-3.1-8B architecture, this model undergoes a unique three-stage training process to boost its cognitive abilities.

Key Capabilities & Innovations

  • Reflective Reasoning Training: Utilizes a proprietary multi-agent system to generate high-quality data for long-thinking tasks, followed by continuous pre-training and supervised fine-tuning.
  • Reinforcement Learning for Reasoning: Incorporates the Skywork o1 Process Reward Model (PRM) to enhance step-by-step reasoning, effectively capturing the influence of intermediate steps on final outcomes.
  • Reasoning Planning: Deploys a proprietary Q* online reasoning algorithm for model-based thinking and searching for optimal reasoning paths, marking its first public implementation.
  • Advanced Cognitive Functions: Exhibits enhanced thinking, planning, self-reflection, and self-verification capabilities.
  • Benchmark Performance: Shows notable improvements across various mathematical and coding benchmarks, outperforming prior models of similar size like Qwen-2.5-7B instruct in its category.

Ideal Use Cases

  • Complex Problem Solving: Adept at handling common-sense, logical, mathematical, ethical decision-making, and logical trap problems.
  • Code Generation & Analysis: Demonstrates strong performance in coding benchmarks.
  • Educational Tools: Can be used for applications requiring detailed, step-by-step reasoning and explanations.
  • Research & Development: Suitable for exploring advanced reasoning and planning in AI models.

Popular Sampler Settings

Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.

temperature
–
top_p
–
top_k
–
frequency_penalty
–
presence_penalty
–
repetition_penalty
–
min_p
–