Chancellor-VLM 40B
Multimodal: text + images, openly
Not released yet — every artifact below ships on launch day. Until then, the Actionfield preview simulates the experience.
About this model
Chancellor-VLM reads figures, lab plots, documents, and photographs alongside text. It was built for science: extracting structure from charts, reasoning about experimental imagery, and reading the tables in papers.
Its vision corpus is the part we're proudest of. Multimodal training data is usually scraped first and questioned later; ours carries the same full audit as our language corpora — provenance and consent documentation for every image source. It took a year longer to assemble. It was worth it.
Architecturally, the model couples a vision encoder to the Chancellor action backbone with a fusion layer trained to preserve spatial detail — which is why it can tell you not just that a curve rises, but where the knee is and what the axis labels say. Up to eight images share a 64K-token context, so it can compare figures across a whole paper.
When text and visual evidence conflict, Chancellor-VLM is trained to flag the disagreement and state which source it weighted. In evaluation, that single behavior eliminated most silent hallucinations on chart-reading tasks.
What it’s good at
- →Chart, plot, and table extraction from papers
- →Visual reasoning over experimental imagery
- →Document OCR with layout understanding
- →Flags text-image disagreement instead of guessing
- →Consent-documented, fully audited vision corpus
Where people use it
How it was trained — in the open
Consent-first corpus
Every image source carries provenance and consent documentation — audited and published, a first at this scale.
Vision-language fusion
A spatial-detail-preserving fusion layer trained on scientific figures, not just web photos.
Grounded instruction tuning
Chart-reading demonstrations paired with the underlying data tables, so answers ground in values, not vibes.
Disagreement training
Deliberately conflicting text-image pairs teach the model to surface contradictions rather than smooth over them.
Benchmarks, with context
Numbers without baselines are marketing. Every score below ships with its comparison point and its caveat.
| Benchmark | This model | Reference | Note |
|---|---|---|---|
| SciChart-QA (chart reading) | 78.9 | 74.2 · best open VLM peer | grounded in axis values |
| DocTable extraction (F1) | 91.3 | 88.0 · best open VLM peer | layout-aware OCR |
| LiveEval-V (contamination-resistant) | 58.4 | 53.9 · best open VLM peer | images postdate cutoff |
| Silent hallucination rate | 3.1% | 11.7% · open VLM average | lower is better |
Everything ships at launch
Launches November 11, 2026. Want a note the moment they’re live? Get notified.