Unverified50% confidenceFactExact time
The runnable configuration points documented by Olmo-core 3 range from a 12.9-billion-parameter model on 16 GPUs all the way to a 1.2-trillion-parameter model on 512 GPUs, and come with official pretraining scripts and a checkpoint conversion tool for exporting to Hugging Face format.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/14/2026
First Seen
Valid until: 1/12/2027
Sources
Related Entities
Related Claims
UnverifiedOlmo-core 3 keeps routing metadata on the GPU, eliminating the host round-trips the old path made to split sizes for all-to-all and group sizes for expert matrix multiplications; device-side scheduled grouped GEMM is used to handle the uneven per-expert shapes produced by routing.74% similarUnverifiedIn the 1.2-trillion-parameter configuration, combined with MXFP8 and per-layer recomputation, Olmo-core 3 reached 858 useful-model TFLOP/s per GPU.73% similarUnverifiedMulti-GPU training supports automatic device allocation, enabling larger models to be split across multiple graphics cards.70% similarUnverifiedOllama's OLLAMA_MAX_LOADED_MODELS default value is 3 times the number of GPUs (3 for CPU inference)70% similarUnverifiedGrok 4.5 is based on the all-new V9 base model, with parameters reaching 1.5 trillion—three times those of the previous-generation V8—and was trained using tens of thousands of NVIDIA GB300 GPUs.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1025295API
curl https://kongchang.com/api/v1/knowledge/claims/1025295MCP
get_claim(id=1025295)