Chancellor-1 8B Instruct
Fast and light — runs anywhere
Not released yet — every artifact below ships on launch day. Until then, the Actionfield preview simulates the experience.
About this model
Chancellor-1 8B is an Action Learning Model that distills the flagship's planning and execution policies into a package that runs on a single consumer GPU — or a laptop. It shares the same fully documented training data and the same open evaluation suite.
Distillation was done transparently: the 8B learned from its 70B sibling's published outputs on the open corpus, and the distillation set is itself released. Nothing about how the small model got its abilities is a secret.
The engineering priority was the first token, not the last benchmark point. On a mid-range consumer GPU the 8B responds in under a second, and quantized GGUF builds bring it to laptops and single-board computers at 6 GB of memory. It is the model we actually deploy — in classrooms, on field robots, and in community projects where a datacenter isn't an option.
It is also the recommended starting point for fine-tuning research: a full LoRA run fits in an afternoon on one GPU, and our released recipes reproduce our own fine-tunes exactly.
What it’s good at
- →Runs locally on 12 GB of VRAM (quantized: 6 GB)
- →Sub-second first-token latency on consumer hardware
- →Same open data audit as the 70B flagship
- →Transparent distillation — the teacher data is public
- →Ideal for fine-tuning experiments and education
Where people use it
How it was trained — in the open
Shared corpus
Same audited 4.1T-token recipe as the flagship, subsampled with a published curriculum for the smaller capacity.
Open distillation
Learned from released Chancellor-1 70B outputs; the full distillation set is downloadable.
Instruction tuning
The flagship's SFT set, filtered for tasks where an 8B can realistically excel — honesty over pretension.
Quantization QA
Every released GGUF build re-runs the full eval suite; degradation per bit-width is published, not hidden.
Benchmarks, with context
Numbers without baselines are marketing. Every score below ships with its comparison point and its caveat.
| Benchmark | This model | Reference | Note |
|---|---|---|---|
| GradSci-QA (graduate science) | 54.6 | 52.3 · best open 8B peer | honest: use the 70B for frontier work |
| LiveEval (contamination-resistant) | 49.1 | 44.7 · best open 8B peer | items postdate training cutoff |
| First-token latency (RTX 4070) | 0.4 s | 1.9 s · 70B on same GPU* | *70B offloaded; 8B fits in VRAM |
| ToolUse-Hard | 66.8 | 61.0 · best open 8B peer | open tool schemas |
Everything ships at launch
Launches October 10, 2026. Want a note the moment they’re live? Get notified.