vectionlabs/Salience-27B-R4

Overview

Salience 27B is a 27-billion-parameter dense vision-language model developed by Vection Labs, designed for demanding software engineering tasks. Unlike MoE models, Salience 27B activates all 27 billion parameters on every token, providing maximum per-token capacity for complex problems. It integrates a hybrid linear+full attention stack for efficient long-context processing and an MTP head for self-speculative decoding, enabling faster inference for its size.

Key Capabilities

• Dense Capacity: All 27B parameters are active per token, ensuring deep processing for challenging problems.
• SWE-Agent First: Optimized for generating runnable code, performing repo-scale edits, methodical debugging, and well-formed native tool calls.
• Terminal & Agentic Work: Capable of planning command sequences, checking results, and recovering from failures in terminal environments.
• Efficient Reasoning: Thinks natively before answering, providing concise and high-quality responses without needing explicit "think step by step" prompts.
• Genuinely Multimodal: Processes images and video as first-class inputs, allowing for understanding diagrams, UI screenshots, stack traces, and whiteboard photos.
• Extended Context: Features a native 262,144-token context window, extendable up to 1,048,576 tokens using YaRN.

Good For

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