An open-weight family for real engineering work — native reasoning, terminal agency, and vision. Read the repo. Plan the commands. Land the patch. Released under Apache-2.0, always.
Salience models are 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. Thinking is native, never performed. Vision is first-class, never bolted on.
Every parameter on, every token — and it stops thinking once it has the answer. Most reasoning models spend the same tokens on what does this flag do as on why does this deadlock under load. R5 reasons in proportion to difficulty, and that is the default rather than a setting you have to find.
The software-engineering agent. 256 experts, ~3B active per token — the decode speed of a small model with the reach of a large one. Repo-scale edits, native XML tool calls, hybrid attention for long horizons.
Fast multimodal for long-horizon work — image and video understanding at sparse-expert speed, tuned for tasks that run for hours rather than seconds.
The 9B that finishes the job. Agentic tool use and native thinking on a single consumer GPU — the tier you actually run locally.
Where the family started. The original 9B — still open, still downloaded hundreds of thousands of times, still doing the job it was built for.
Thinking-native and frontend-obsessed — software engineering and interface work in a 9B body. A gift to the community that got us here.
Every release ships full weights under Apache-2.0. No gates, no waitlists, no research-only licenses. If you can download it, you own your copy.
Our models reason inside <think> because they were built to — never because a prompt begged them to. You budget the tokens; they do the thinking.
Well-formed tool calls, planned command sequences, step-by-step verification. A Salience model checks each result before taking the next step — and recovers from failure instead of repeating it.
Screenshots, diagrams, whiteboard photos, video clips — engineering inputs, not afterthoughts. Debug straight from a stack-trace screenshot, mid-task.
Every model is 🤗 transformers-native. Pull, load, work.
from transformers import AutoModelForImageTextToText, AutoProcessor
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": "Find and fix the off-by-one error in this loop."}]
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])Everything we make is on Hugging Face. Take it. Build with it.
huggingface.co/vectionlabs