vectionlabs/Salience-27B-R5

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 17, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Salience-27B-R5 by Vection Labs is a 27-billion-parameter dense vision-language model built on the Qwen3.8 architecture, featuring a 1,048,576-token context window. It is specifically engineered for hard, practical software engineering tasks, including code generation, debugging, repo-scale edits, and multi-step terminal agency, with native multimodal perception. This model uniquely offers 'reasoning economy' by default, proportionally adjusting its deliberation based on problem difficulty, and utilizes a hybrid linear+full attention stack for long-context speed.

Loading preview...

Salience-27B-R5: A Dense Vision-Language Engineer

Salience-27B-R5, developed by Vection Labs, is a 27-billion-parameter dense vision-language model built on the Qwen3.8 architecture. It is specifically designed for hard, practical engineering work, excelling in tasks like writing and debugging code, performing repo-scale edits, enabling multi-step terminal agency, and quantitative reasoning. A key feature is its native vision capability and an extensive 1,048,576-token context window.

Key Differentiators

  • Reasoning Economy by Default: Unlike many models that apply the same reasoning effort to all problems, Salience-27B-R5 proportionally adjusts its deliberation based on problem difficulty. It reasons deeply when needed and provides direct answers otherwise, making its thinking process more efficient.
  • Dense Capacity: All 27 billion parameters are active on every token, ensuring maximum depth for complex problems without the routing overhead of MoE models.
  • SWE-Agent First: Tuned to produce runnable code, facilitate repo-scale edits, enable methodical debugging, and generate well-formed native tool calls.
  • Multimodal Perception: Genuinely multimodal, accepting images and video as first-class inputs for tasks like reading diagrams, UI screenshots, or stack-trace photos mid-task.
  • Fast Decode: Incorporates a hybrid linear+full attention mechanism and an MTP head for self-speculative decoding, offering efficient performance for its size.

Intended Use Cases

  • Code generation, explanation, debugging, and review.
  • Terminal and tool-using agent workflows (CLI agents, browsing, ML engineering).
  • Backend and systems design, infrastructure-as-code.
  • Step-by-step reasoning and quantitative problem solving.
  • Understanding diagrams, screenshots, and documents within engineering contexts.

This model is released under the Apache-2.0 license and is transformers-native, supporting AutoModelForImageTextToText.