Unverified50% confidenceFactExact time
Beam's reinforcement learning phase used 1,500 GB300 chips running continuously for four weeks, generating over 100 million rollouts, with a maximum context length of 256K.
1
Sources
50%
Confidence
Long-term
Relevance
10/8/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedBeam's pre-training used 6,144 NVIDIA GB300 NVL72 chips, took less than 4 weeks, and consumed 23.8 trillion tokens.76% similarUnverifiedA tester ran a 20.7B dense model on a dual-4090 (48 GB VRAM) workstation and achieved 27.6 tokens per second, only slightly faster than Dwarf Star, but with over 10 times fewer parameters.68% similarUnverified开发者公布在RTX 3050 Laptop（4GB）上微调Llama-3.1-8B的实测结果为训练速度119.6 tok/s，峰值显存3.32 GB66% similarUnverifiedThe program recommends context length based on VRAM: 8 GB VRAM is recommended for a 32K context, and 24 GB VRAM can support 128K.65% similarUnverifiedTaking a 70B BF16 model as an example, generating one token requires reading approximately 140 GB of weights, and when fully utilizing the H100's 3.35 TB/s memory bandwidth, the per-token latency is at least 42 milliseconds, with a theoretical TPS upper limit of approximately 24 tokens/second.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/987287API
curl https://kongchang.com/api/v1/knowledge/claims/987287MCP
get_claim(id=987287)