Unverified50% confidenceFactExact time
REMORY converges after 250,000 training steps on 64 A100s; the final released checkpoint has approximately 350 million parameters and occupies 1.4GB of GPU memory.
1
Sources
50%
Confidence
Long-term
Relevance
10/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe TEMPEST model has only 720,000 parameters, a storage footprint of 2.80 MB, and is able to converge within 50 training epochs.75% similarUnverifiedOn an A100 (80 GB VRAM) GPU, the throughput for training a 7B model is approximately 3,000–5,000 tokens per second.75% similarUnverifiedThe demo constructed a Float32Array of scale 10000×10000×10 as test data, reaching GB-level size, with a single computation taking over one second, of which approximately one second was the initialization of 4 GB of memory space rather than language interop overhead.70% similarUnverifiedDuring training, the Adam optimizer additionally stores gradient momentum at approximately 2–3 times the weight size, meaning the full training memory requirement for a 70B model can easily exceed 400 GB.70% similarVerifiedFor a typical 7B model with 32 layers and 32 attention heads, storing the KV Cache for each token requires approximately 0.5 MB of VRAM, so generating 1,000 tokens requires an additional ~500 MB of VRAM69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/998485API
curl https://kongchang.com/api/v1/knowledge/claims/998485MCP
get_claim(id=998485)