You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Thinking Machines have just released Inkling-2-Small. It has roughly the same amount of parameters/active parameters as DeepSeek V4.1, and I think it would be nice to have optimized inference for this model here.
Phase 1 for this is up as a draft in #981: a converter to llama.cpp's inkling2 layout (the same format as unsloth's GGUFs), a reference forward pass, and tests. The real model matches transformers (KL 1e-11). There's no engine yet. Happy to do the kernels next, or leave it as a base if you'd rather build it.
Activity
mrsirg97-rgb commentedon Nov 13, 2026
Phase 1 for this is up as a draft in #981: a converter to llama.cpp's
inkling2layout (the same format as unsloth's GGUFs), a reference forward pass, and tests. The real model matches transformers (KL 1e-11). There's no engine yet. Happy to do the kernels next, or leave it as a base if you'd rather build it.