Cassandra T1
1.33B masked-diffusion model. Weights, checkpoint and SHA-256 are public under Apache 2.0.
Open →Seven stops. Four are finished, two are being built, and one is waiting on funding. We do not print dates we cannot keep, so the order is the plan.
Public, and checkable today.
1.33B masked-diffusion model. Weights, checkpoint and SHA-256 are public under Apache 2.0.
Open →147.5M parameters, 35 specialists. Anchor-token masking beats random masking by 1.67× held out, with a DOI and 504 published generations.
Open →A 360M compositional edge model with sealed specialists, running in the browser.
Open →A website platform for language models to run: one Rust binary, 53 MCP tools, signed releases. Live with a demo.
Open →Being built on our own hardware.
430M masked-diffusion MoE: one shared expert and 16 routed experts, top-2. The first step to Sophia-4B.
Open →Sophia Reader-Writer, 431M parameters: a 16-layer MoE reader and a dense writer. 291 MB on a phone.
Open →Planned, with a budget and written pass criteria.
3.88B total, 1.24B active, 256K context. Open weights and a hosted API. About $200K on rented H100s.
Open →S0 shows the architecture at a size we can train and test quickly. SRW Pocket tests it on a phone with no network. Sophia-4B is the same design at 3.88B. Pass criteria are written before each run and the result is published whichever way it goes.