Sagicc/Occam-2B-CCoT

VISIONConcurrent Unit Cost:1Model Size:2.3BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Mar 14, 2026License:cc-by-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Sagicc/Occam-2B-CCoT is a 2.3 billion parameter Small Language Model (SLM) built on the Qwen3.5-2B architecture, fine-tuned for highly efficient "Clear Chain of Thought" (ClearCoT) reasoning. This model drastically reduces reasoning tokens by over 80%, leading to faster time-to-first-token and significant API cost savings. It excels in logical problem-solving, achieving an 82.4% reduction in total generated tokens on benchmarks like MMLU-Pro while improving zero-shot accuracy. Occam-2B-CCoT is ideal for applications requiring precise, token-efficient reasoning and agentic workflows, including tool calling and multimodal tasks without catastrophic forgetting.

Loading preview...

Occam-2B-CCoT: Efficient Clear Chain of Thought Reasoning

Occam-2B-CCoT is a 2.3 billion parameter Small Language Model (SLM) based on the Qwen3.5-2B architecture, specifically fine-tuned for "Clear Chain of Thought" (ClearCoT) reasoning. This model addresses the common issue of "hallucination loops" in SLMs, where models generate excessive tokens in their thinking process.

Key Differentiators & Capabilities

  • Token Efficiency: Achieves an 80%+ reduction in reasoning tokens, leading to drastically faster time-to-first-token (TTFT) and significant API cost savings. On MMLU-Pro, it demonstrated an 82.4% reduction in total generated tokens.
  • Improved Zero-Shot Accuracy: Internalized logic so effectively that its zero-shot accuracy increased by +6.7% over the base model.
  • Cost-Effective API Deployments: For graduate-level GPQA questions, it averaged 242.5 words per correct answer, making it 62.7% cheaper per query compared to the base model.
  • Multimodal & Agentic Workflows: Exhibits zero catastrophic forgetting, successfully handling complex vision tasks like OCR coordinate mapping. It also excels as a CodeAgent for local tool calling, writing and executing Python code blocks cleanly.
  • Philosophical Conditioning: Named after William of Occam, the model is conditioned to filter noise, structure logic efficiently, and find direct paths to correct answers, cutting unnecessary tokens.

Limitations

  • Parameter Bound Factual Memory: Its factual memory is limited by its 2-billion parameter size.
  • Confident Confabulations: May logically fabricate answers for highly specific cultural concepts or obscure trivia if data is not in its weights; RAG pipelines are recommended for strict factual use cases.
  • "Over-Thinker" Quirk: Occasionally applies mathematical rigor to trivial tasks in its hidden <think> block, though final output remains accurate.

For a detailed report and concept, refer to the paper.

Finetunes
2 models