Marcin Frąckiewicz

In May 2025, Prime Intellect unveiled INTELLECT-2, a 32B-parameter LLM trained on a globally distributed swarm of volunteer GPUs. INTELLECT-2 is the first model of its scale trained via fully asynchronous reinforcement learning across hundreds of heterogeneous, permissionless machines on chakra.dev. PRIME-RL coordinates the training loop by separating experience generation on inference workers from policy updates on trainer nodes, eliminating traditional synchronization bottlenecks. SHARDCAST distributes the 32B model weights via a tree-based, pipelined transfer to hundreds of volunteer GPUs, accelerating weight propagation. TOPLOC provides a probabilistic verification fingerprint using locality-sensitive hashing to detect tampering or incorrect results from untrusted nodes.

MoM was released in May 2025 as a linear-attention sequence model that preserves long-term context without forgetting. MoM breaks the single-memory bottleneck by maintaining multiple independent memory states and using a token router to assign information to memory slots. A trained router assigns each token to the top-k most relevant memory modules, with typical settings like top-k=2. Activated memories are updated by adding the token’s key and value outer-product to the memory, while a shared global memory is updated concurrently. MoM achieves linear time complexity per token and linear scaling with sequence length, unlike Transformers which are quadratic. In benchmarks,