Research terminal

AirLLM lowers peak accelerator memory by loading model layers sequentially instead of keeping the full checkpoint resident during inference.

Cache-to-Cache learns projections that fuse a source model's KV cache into a receiver, letting paired language models exchange latent state without generating intermediate text.

Dia is a 1.6-billion-parameter model for generating multi-speaker dialogue from tagged transcripts, with optional audio conditioning and nonverbal cues.

DigUp builds an on-device multimodal index for files and retrieves photos, documents, recordings, video moments, and code from natural-language or exact-term queries.

FastAPI-MCP exposes selected FastAPI operations as MCP tools while reusing request schemas, endpoint documentation, ASGI integration, and existing dependency-based authentication.

Handwired compiles a constrained JavaScript program into matrix operations and neural-network weights, producing an exact differentiable network without a training step.

KV Cache Transform Coding applies learned decorrelation, adaptive quantization, and entropy coding to reduce storage and transfer costs for reusable language-model KV caches.

LangExtract turns unstructured text into schema-shaped records while mapping extracted values back to character spans and generating a reviewable visualization.

OmniExtractBench provides a common structured-document extraction dataset, runner, and deterministic scorer whose verdicts distinguish matches, omissions, misreads, and fabricated values.

ZeroSearch trains a language model to use search-like reasoning while another model simulates retrieval results, removing live search API calls from reinforcement-learning rollouts.