Open source · Open models

Models with nothing to hide

The Chancellor family of Large Action Models is open the way science is supposed to be: weights, training data, code, and evaluations, all under Apache-2.0. Launching one a month this autumn — September 9, October 10, and November 11, 2026. Nothing is downloadable yet; everything will be.

Try them in the Actionfield→Compare models

The standard

What “fully open” means here

Open weights

Every parameter, every checkpoint — including the intermediate ones most labs delete.

Open data

The full training recipe with a token-level audit. You can see exactly what each model read.

Open code

Training, fine-tuning, and serving code under Apache-2.0. Retrain from scratch if you want to.

Open evals

Contamination-resistant benchmarks, published with every result — the good ones and the bad ones.

The family

Three models, one standard

Flagship · 70B

Chancellor-1 70B Instruct

Chancellor-1 is the strongest fully open Large Action Model (LAM) we know how to build. It doesn't just answer — it plans and executes multi-step actions with tools. Seventy billion parameters trained on 4.1 trillion fully documented tokens, it matches closed frontier systems on graduate-level science questions — and unlike them, it can prove where its knowledge comes from.

That proof matters. Because the corpus is public, Chancellor-1 supports ChancellorTrace: for any response, you can see documents from the training data with exact text matches. Capabilities stop being magic and become evidence.

→128K-token context for full-paper and multi-document analysis
→Tool calling with open, auditable tool schemas
→Every intermediate checkpoint published — study how ability emerges

Launches Sep 9, 202670B params4.1T tokens128K context

Model page→Try it live

Efficient · 8B

Chancellor-1 8B Instruct

Chancellor-1 8B is the flagship's reasoning distilled into a model that runs on a single consumer GPU — or a decent laptop. Same documented data, same open evaluation suite, a fraction of the footprint.

It exists because openness only matters if people can actually run the model. The 8B is what we deploy in classrooms, field robotics, and community projects: fast enough for real-time use, small enough to fine-tune on a weekend, and honest enough to tell you when a question deserves its 70B sibling.

→Runs on 12 GB VRAM — quantized builds fit in 6 GB
→Sub-second first token on consumer hardware
→LoRA and full fine-tuning recipes included

Launches Oct 10, 20268B params64K contextGGUF builds

Model page→Try it live

Multimodal · 40B

Chancellor-VLM 40B

Chancellor-VLM reads what science actually looks like: figures, lab plots, tables, microscopy, field photographs. Built on the Chancellor action backbone, it adds a vision system trained on the first fully audited multimodal corpus — provenance and consent documented for every image source.

Ask it to pull the numbers out of a chart, compare two experimental images, or read a scanned table into clean data. Where text and visual evidence disagree, it says so — and tells you which one it weighted, and why.

→Chart, plot, and table extraction from papers
→Up to 8 images in a 64K-token context
→Consent-documented vision corpus — a first at this scale

Launches Nov 11, 202640B params8 images64K context

Model page→Try it live

Side by side

Pick your Chancellor

	Chancellor-1 70B	Chancellor-1 8B	Chancellor-VLM
Availability	Sep 9, 2026	Oct 10, 2026	Nov 11, 2026
Parameters	70B	8B	40B
Modality	Text	Text	Text + images
Context window	128K tokens	64K tokens	64K tokens + 8 images
Runs on	4× datacenter GPUs	1 consumer GPU / laptop	2× datacenter GPUs
Tool calling	Yes	Yes	—
Best for	Frontier reasoning, research	Local apps, education, fine-tuning	Charts, documents, lab imagery
License	Apache-2.0	Apache-2.0	Apache-2.0

One launch a month, all autumn

Chancellor-1 70B on September 9, the 8B on October 10, and Chancellor-VLM on November 11. Every artifact ships on launch day — and if you build something better from our recipe, that’s not competition, that’s the point.

Get notified at launch→Preview the flagship