Unverified50% confidenceFactExact time
RadixArk/Qwen3.8-Flash-Next-NVFP4 mentioned in the LMSYS blog is an SGLang checkpoint that uses post-training quantization, quantizing only the routed experts, calibrated with 128 CNN/DailyMail articles truncated to 512 tokens, validated on GB300 and B300, and reported to reach 540 tok/s at batch size 1 with B200 TP4 plus MTP.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/17/2026
First Seen
Valid until: 10/31/2026
Sources
Related Entities
Related Claims
UnverifiedThe dataset is hosted on Hugging Face (v0rt3ch/cs-5v5-bots-aim_usp-3k), and the publisher provides full specifications at vor.tech.64% similarUnverifiedThe NVFP4 version of Qwen3-27B uses sensitivity-aware mixed precision: feed-forward layers run in FP4 (E2M1) with FP8 block scaling, while attention projections, DeltaNet recurrent layers, the vision encoder, and multi-token prediction heads remain in FP8 or BF16.63% similarUnverifiedQwen3.8-Flash-Next uses a hybrid attention mechanism combining GDN and QSA, with the MoE layers using 512 experts and Top-10 routing.63% similarUnverifiedIn the experiment, the Qwen3.5-0.8B backbone and approximately 51 billion parameter PLE memory module were frozen, with only R=1 readers deployed at decoder layers 3 and 9 being trained.62% similarUnverifiedQwen3.8-Flash-Next adopts a hybrid attention mechanism of GDN plus QSA, with a total of 125B parameters plus 51B N-gram embeddings, activating only approximately 6B parameters per token.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1049822API
curl https://kongchang.com/api/v1/knowledge/claims/1049822MCP
get_claim(id=1049822)