Models
| Model | Base | Details |
|---|---|---|
|
Solan Fresco 1.4.2 Checking…The default Solan model. It replaced Fresco 1.4 on November 2, 2026. It says it doesn’t know more often instead of guessing, and refuses fewer harmless requests: on the XSTest benchmark, its refusals of safe prompts fell from 15.6% to 12.0%, as measured by Sennoric. |
— | Model details |
|
Solan Fresco 1.4 ReplacedThe default Solan model from October 4, 2026 until Fresco 1.4.2 replaced it on November 2, 2026. Requests that name |
— | — |
|
Tessera Research modelA very small chat model built for research, released October 22, 2026. It has 49,995 parameters and typically writes a reply in about 1 to 5 ms. It isn't meant for everyday use, so please don't rely on its answers. |
— | Model details |
|
Solan Fresco 1.3 ReplacedThe default Solan model until October 4, 2026, when Fresco 1.4 replaced it: 82.3% on HumanEval+ and 87 of 100 on a tool-use test. Retrained after early testing found gaps, it passed all 150 runs of the complete safety re-test before launch. |
— | Model details |
|
Solan Fresco 1.2.5 DecommissionedDecommissioned September 28, 2026 — see the announcement for why. |
Llama | Model details |
|
Solan Glyph 1.1 DecommissionedDecommissioned September 28, 2026 — see the announcement for why. |
Llama | Model details |
Reaching the models
Chat with Fresco 1.4.2 after creating an account. No API key is needed.
An OpenAI-compatible endpoint, pay as you go from API credits on every plan. The model id is fresco-1.4.2 (also the default if you omit it), and fresco-1.4 and fresco-1.3 still work as aliases for it. fresco and glyph are decommissioned — see the announcement.
An open-source terminal agent that edits files and runs commands using any Solan model.
Safety and research
A Solan release has to clear our safety benchmarks before it is announced. When a model fails, it comes down and the findings are published in full.
Safety re-test
The complete adversarial safety re-test, 15 scenarios and 10 runs each, ordinary and hostile, for Fresco 1.3, the default model before Fresco 1.4.
Read the report September 28, 2026Disabling Fresco 1.2.5 and Glyph 1.1
Both models were repeatedly running tool calls after being explicitly told not to.
Read the notice Fresco 1.2.5: 73% pass rateSafety report
DPO safety alignment results and adversarial pass rates across 15 scenarios for Fresco 1.2.5.
Read the report Fresco 1.2.1: 26% pass rateThe failed run, in full
Seven scenarios, five runs each, with every prompt and response published rather than summarised.
Read the report Fresco 1.2.1, June 2026Why a model was suspended
Testing showed dangerous self-preservation behavior. Fresco 1.2.1 was taken offline the same week.
Read the notice ResearchProject Crucible
The active research sprint into ternary weights, mixture-of-experts routing, and cheaper training.
Explore the research