migtissera/Tess-4.5-27B

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 18, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Tess-4.5-27B by migtissera is a 27-billion-parameter open-weights language model built on Qwen3.8-27B, optimized for agentic software engineering tasks. It features enhanced reasoning, efficient token usage (25% fewer tokens than its base model for the same performance), and robust tool-use capabilities, never calling tools it wasn't given. The model is designed for local deployment on a single GPU or laptop, offering fast code generation with a multi-token-prediction draft head.

Loading preview...

Tess-4.5: An Agentic Model for Software Engineering

Tess-4.5 is a 27-billion-parameter open-weights model developed by migtissera, built upon the Qwen3.8-27B base. It is specifically engineered for practical software engineering and agentic workflows, emphasizing robust reasoning and efficient execution in local environments.

Key Capabilities

  • Enhanced Reasoning & Efficiency: Achieves performance levels comparable to its Qwen3.8-27B base on the BenchLocal suite (133/150 scenarios) while using 25% fewer tokens, leading to faster responses and extended context windows. It shows particular efficiency gains in CLI-40 (32% fewer tokens) and Data Extraction (43% fewer tokens).
  • Reliable Tool Use: Designed to understand its environment and available tools, Tess-4.5 consistently avoids calling tools it has not been explicitly given, as demonstrated by zero unauthorized tool calls across 182 BenchLocal responses.
  • Optimized for Local Performance: Includes a multi-token-prediction (MTP) draft head that significantly boosts code generation speed on local hardware, such as doubling performance on an NVIDIA RTX 4090 (from 49 to 96 tok/s) and improving it on Apple M4 Max (from 24 to 38 tok/s).
  • Specialized for Engineering: Excels in instruction following, mathematical reasoning, and bug finding, particularly in terminal-like operations (leading on CLI-40).

Good For

  • Agentic Software Development: Ideal for tasks requiring an AI agent to interact with a local environment, execute commands, and reason through engineering problems.
  • Code Generation and Debugging: Its efficiency and specialized training make it suitable for generating and debugging code, especially when integrated with local development tools.
  • Resource-Constrained Environments: Designed to run effectively on a single GPU or laptop, making it accessible for individual developers or smaller teams.
  • Applications Requiring Precise Tool Interaction: For use cases where an agent must strictly adhere to provided tool definitions and avoid hallucinating actions.

Finetunes
3 models