sabicap/sabi-noema

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 13, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Sabi-Noema is a 27 billion parameter agentic model developed by sabicap, built upon Qwen3.8-27B and fine-tuned using open-source RL environments. It excels in agentic tasks, achieving 76.0% on Terminal-Bench 2.1 and significantly reducing hallucination rates to 15.1% compared to its base model. This model is optimized for reasoning efficiency and long, multi-turn agentic jobs, demonstrating 2.3x fewer output tokens than Qwen3.8-27B on Terminal-Bench tasks.

Loading preview...

Sabi-Noema: An Efficient Agentic Model

Sabi-Noema is a 27 billion parameter agentic model developed by sabicap, built on the Qwen3.8-27B architecture. It has been exclusively trained using open-source Reinforcement Learning (RL) environments, focusing on enhancing agentic capabilities and reasoning efficiency while reducing output token usage.

Key Capabilities

  • Superior Agentic Performance: Achieves 76.0% on Terminal-Bench 2.1 and 47.5% on DeepSWE, outperforming its base model Qwen3.8-27B.
  • Reduced Hallucination: Demonstrates a significantly lower hallucination rate of 15.1% on AA-Omniscience, less than half of the base model's rate.
  • Token Efficiency: Uses 2.3x fewer output tokens than Qwen3.8-27B on Terminal-Bench tasks and 29% fewer on DeepSWE, leading to cost-sensitive serving.
  • Robust Tool Calling: Scores 63.9% on BFCL v4 multi-turn and 48 on MCPMark, indicating strong tool integration.
  • High Context Length: Supports a context length of 262K tokens, suitable for complex, long-running tasks.

Good For

  • Long, Multi-Turn Agentic Jobs: Ideal for tasks requiring many actions and checks, such as shell operations or tool-based workflows.
  • Customer-Facing Agents: Suitable for policy-bound agents, including those handling difficult customer interactions, due to its reduced hallucination and correct tool calls.
  • Cost-Sensitive Deployments: Its token efficiency makes it a strong candidate for applications where output token costs are a concern.
  • Assistants Requiring Reliability: Useful for applications where the model needs to explicitly state when it lacks information, thanks to its low hallucination rate.
  • High-Density Serving: The related Sabi-Noema-MLA variant is designed for many long sessions on a single GPU by shrinking the attention cache.

Finetunes
2 models