vectionlabs/Salience-27B-R6

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 6, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Vection Labs' Salience-27B-R6 is a 27-billion-parameter dense vision-language model built on the Qwen3.8 architecture, featuring native vision and a 1,048,576-token context window. It runs all 27B parameters on every token, utilizing a hybrid linear + full attention stack and an MTP head for self-speculative decoding. This model is specifically engineered for practical software development tasks, including code generation, debugging, repo-scale edits, and multi-step terminal agency, with a focus on 'reasoning economy' to provide proportional deliberation.

Loading preview...

Salience-27B-R6: A Dense Vision-Language Engineer

Salience-27B-R6 by Vection Labs is a 27-billion-parameter dense vision-language model, the sixth revision in its tier, designed for demanding engineering tasks. Unlike Mixture-of-Experts (MoE) models, Salience-27B-R6 activates all 27 billion parameters on every token, ensuring maximum per-token capacity. It features a hybrid linear + full attention stack for long-context speed and an MTP head for self-speculative decoding, enhancing decode speed.

Key Differentiators & Capabilities

  • Reasoning Economy: The model's defining property is its ability to reason proportionally to the problem's difficulty, stopping once it has the answer. This is now a default behavior, leading to shorter reasoning chains for the same answer compared to previous versions.
  • Dense Capacity: All 27B parameters are active on every token, providing deep processing for complex problems.
  • Vision-Language Integration: Genuinely multimodal, accepting images and video as first-class inputs for tasks like reading diagrams, UI screenshots, or stack traces.
  • Extended Context Window: Offers a 1,048,576-token context window, enabling analysis of entire repositories or extensive documentation.
  • Optimized for Software Engineering: Tuned for generating runnable code, performing repo-scale edits, methodical debugging, and robust tool calls.
  • Terminal & Agentic Work: Excels at multi-step planning, tool orchestration, and long-horizon execution within terminal environments.
  • Fast Decode: Achieves fast decoding for its size through a hybrid attention mechanism and an MTP head.
  • Reduced Refusal: Designed to answer questions directly with reduced refusal behavior, without built-in content filters.

Good For

  • Code Generation & Debugging: Creating, explaining, debugging, and reviewing code, including repo-scale tasks.
  • Agentic Workflows: Powering CLI agents, browsing, ML engineering, and DevOps tasks requiring multi-step planning and tool use.
  • Technical Research: Backend and systems design, infrastructure-as-code, and quantitative problem-solving.
  • Multimodal Engineering: Understanding diagrams, screenshots, and documents within engineering workflows.

This model is released with open weights under the Apache-2.0 license and is transformers-native. While optimized for efficiency in agent loops, users should note that on saturated multiple-choice knowledge benchmarks, it may perform slightly behind its R5 predecessor due to its focus on shorter reasoning chains.