Unverified50% confidenceBenchmarkExact time
Comparing BF16 and NVFP4 under the same harness, WebVoyager dropped from 77.72 to 66.31, DeepSWE from 24.65 to 18.2, and MMLU-ProX from 76.91 to 73.76, with other rows within the noise range.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/16/2026
First Seen
Valid until: 1/14/2027
Sources
Related Entities
Related Claims
UnverifiedOn the DeepSWE coding leaderboard, Beam scored 44.4 points, ranking last, below DeepSeek V4.1 Flash's 74.2 points, Kimi K3's approximately 68 points, GLM's approximately 61 points, and Qwen3 Max's approximately 51 points.68% similarVerifiedThe NVIDIA H20's peak BF16 compute is approximately 296 TFLOPS, significantly reduced compared to the H100's 989 TFLOPS, but it retains 141 GB HBM3 memory and 4.0 TB/s memory bandwidth.67% similarUnverifiedCompressing DeepSeek Coder to 2-bit loses 67% of passing solutions, whereas the 4-bit version sacrifices only 2% for a 70% reduction in size.64% similarUnverifiedOn a single RTX PRO 6000, peak VRAM for Qwen3.8-27B-NVFP4 dropped from 72.9 GB to 40.2 GB, and per-step time was reduced by approximately 11%63% similarUnverifiedOn a ResNet18 training step, deterministic mode incurs an overhead of +16.5% compared to the default under fp32 (17.94 ms vs. 20.90 ms), and +15.6% under bf16 autocast (12.42 ms vs. 14.37 ms).63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1049453API
curl https://kongchang.com/api/v1/knowledge/claims/1049453MCP
get_claim(id=1049453)