Unverified50% confidenceFactExact time
Beam's pre-training used 6,144 NVIDIA GB300 NVL72 chips, took less than 4 weeks, and consumed 23.8 trillion tokens.
1
Sources
50%
Confidence
Long-term
Relevance
10/8/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedBeam'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.76% 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 GB62% 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.60% similarUnverifiedGPT-4-level models can have a Time to First Token (TTFT) of 500–2000ms, while smaller models such as Llama 3.1 8B can achieve a TTFT as low as 100–300ms.59% similarUnverifiedTraining an 800B activated-parameter model requires approximately 50,000 GB300 chips; switching to Huawei's latest chips would require 200,000.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/987286API
curl https://kongchang.com/api/v1/knowledge/claims/987286MCP
get_claim(id=987286)