vectionlabs

Salience-1.5-Pro

README

License: apache-2.0

Abstract

Salience 1.5 Pro is a sparse Mixture-of-Experts vision-language model: 35B total parameters, but only ~3B active per token across 256 experts (8 routed + 1 shared). It decodes at the speed of a small dense model while reasoning with the capacity of a 35B one — built for hard, practical engineering work: writing and debugging real code, repo-scale edits, robust backend systems, and driving tools and agents through long-horizon tasks, with native vision and a 262K-token context window.

It is the flagship tier of the Salience family — engineered for people who care less about chat pleasantries and more about whether the model can do the thing: ship the function, find the bug, call the right tool, land the pull request.

Highlights

Bigger and faster at once. Sparse activation means ~3B of compute for 35B of knowledge — and a hybrid linear+full attention stack keeps long contexts fast.
SWE-agent first. Tuned to produce runnable code, repo-scale edits, methodical debugging, and well-formed native tool calls.
Long-horizon agentic execution. Multi-step planning and tool orchestration that hold up across extended task chains.
Reasoning that shows its work. Structured, inspectable chains of thought — with a one-token switch to turn them off when you want instant answers.
Genuinely multimodal. Images and video are first-class inputs — read a diagram, a UI screenshot, a stack-trace screenshot, or a whiteboard photo mid-task.
Long context. 262,144 tokens native — whole repos and long specifications in a single prompt.
Open weights. Apache-2.0, transformers-native, single-file deployment.

Model overview

Table

Parameters	35B total / ~3B active (256 experts, 8 routed + 1 shared)
Modalities	text, image, video -> text
Context window	262,144 tokens native
Attention	hybrid linear + full attention (full every 4th layer)
Precision	bfloat16 (float32 SSM components preserved)
Architecture	Qwen3.5 MoE (35B-A3B) + native vision encoder
License	Apache-2.0
Library

Multimodal MoE sibling: Salience-1.5-Flash (30B-A3B, vision + video + 1M context).

Capabilities

Salience 1.5 Pro routes only ~3B parameters per token through a 35B expert network. Its capability profile is built around four pillars:

Code & SWE execution - runnable code, repo-scale edits, methodical debugging, robust backends.
Agentic tool-use - multi-step planning, tool orchestration, long-horizon task execution.
Deep reasoning - structured, inspectable chains of thought for hard, multi-step problems.
Multimodal perception - diagrams, screenshots, documents, and video as first-class inputs.

Thinking

Thinking is on by default: the model reasons inside <think>...</think> before answering, and serving stacks expose it as reasoning_content. Pass enable_thinking=False to apply_chat_template for instant direct answers.

Tool calling

The model emits XML-style tool calls (<tool_call><function=...><parameter=...>), parsed natively by vLLM / SGLang tool parsers for this model family. Provide tool schemas via the chat template tools argument.

Intended use

Salience 1.5 Pro targets software engineering, coding agents, and technical research:

Code generation, explanation, debugging, review, and repo-scale tasks.
Agentic / tool-using workflows (terminal agents, browsing, ML engineering).
Backend and systems design.
Step-by-step reasoning and quantitative problem solving.
Screenshot / diagram / document understanding inside engineering workflows.

It is not intended for high-stakes decisions without human review, nor as a source of truth for medical, legal, or financial advice.

Quickstart

python
from transformers import AutoModelForImageTextToText, AutoProcessor
import torch

repo = "vectionlabs/Salience-1.5-Pro"
proc = AutoProcessor.from_pretrained(repo)
model = AutoModelForImageTextToText.from_pretrained(
    repo, dtype="auto", device_map="auto"
)

messages = [{
    "role": "user",
    "content": [{"type": "text", "text": "Implement an LRU cache in Python with O(1) get/put."}],
}]
text = proc.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = proc(text=[text], return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=2048)
print(proc.batch_decode(out[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)[0])

Requires a recent transformers (>= 5.8). Vision works the same way with {"type": "image", "image": ...} content items. Pass enable_thinking=False to apply_chat_template for direct answers.

Quantized GGUF (recommended for local)

This is a sparse MoE (~3B active). Use Q6_K minimum (Q5_K_M acceptable) - Q4 and below corrupt the router and cause loops / degeneration. Dense-model Q4 intuition does not apply here.

Prompting tips

Code: specify language, constraints ("no external libraries"), and the exact I/O contract.
Agentic / tools: pass real tool schemas via the template; the model emits XML-style calls.
Reasoning: thinking is on by default; let it externalize its work on hard problems.
Vision: put the image before the question in the message content.
Sampling: temperature=0.6-0.85, top_p=0.95, top_k=20, presence_penalty=1.1 (lower temperature for precise code edits, higher for long agentic sessions).

Deployment

Single / multi GPU: loads with device_map="auto"; keep dtype="auto" so float32 components stay float32.
Serving: vLLM / SGLang with the model family's reasoning parser (<think> -> reasoning_content) and XML tool-call parser.
Quantized formats: GGUF and other community quantizations are supported (Q6_K+ for MoE).

Limitations & responsible use

Salience 1.5 Pro can be confidently wrong. Verify factual and mathematical claims.
Generated code may be insecure or incorrect - review before running, never execute untrusted output.
Long-context and video inputs increase latency and memory substantially.
Do not use it for surveillance, manipulation, or any use that violates applicable law or the Apache-2.0 terms.
No audio modality.

Citation

bibtex
@misc{vectionlabs2026salience15pro,
  title  = {Salience 1.5 Pro: A Sparse-MoE Software-Engineering Agent},
  author = {Vection Labs},
  year   = {2026},
  url    = {https://huggingface.co/vectionlabs/Salience-1.5-Pro}
}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

vectionlabs

Model Tree

Base

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality