Unverified50% confidenceFactExact time
Running the BF16 checkpoint of Muse Spark Open requires about 2 TB of VRAM, equivalent to a cluster of 32 H200 GPUs; the NVFP4 (4-bit) checkpoint reduces the requirement to about 530 GB, validated on four NVIDIA Blackwell Ultra GPUs or eight H200s.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/16/2026
First Seen
Valid until: 1/14/2027
Sources
Related Entities
Related Claims
UnverifiedThe 816B-parameter Muse Spark Open occupies about 493GB of memory after NVFP4 quantization, while FP16 storage would require more than 1.6TB.77% similarVerifiedRunning Muse Glimmer at full precision requires 55 GB of VRAM, which is compressed to under 20 GB via 4-bit quantization, allowing it to run on consumer-grade GPUs with 24 GB of VRAM.70% similarUnverifiedNVIDIA released an NVFP4 quantized model of Qwen3 235B on Hugging Face, compressing the model from approximately 470 GB of VRAM in BF16 to approximately 156 GB, enabling it to run on a single-node cluster of 8 H100 GPUs.68% similarUnverifiedOn 8 NVIDIA B300 GPUs, a 47-billion-parameter MoE model reaches about 52,000 tokens per second per GPU on the new Olmo-core 3 framework, whereas the old implementation reached only about 19,400, a throughput improvement of about 2.7 times; Ai2 labels this as a "preliminary result."68% similarUnverifiedThe BF16 weight files for Naive-N0.5-Flash are approximately 618GB, and the FP8 format version weights are approximately 315GB67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1047126API
curl https://kongchang.com/api/v1/knowledge/claims/1047126MCP
get_claim(id=1047126)