Unverified50% confidenceFactExact time
NVFP4 quantization is not 4-bit for the whole model: the vision encoder, connector, Mamba convolutions, and embedding stay in BF16, and the MoE router stays in FP32.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedQwen3.8 27B officially provides a BFloat16 full-precision version and an FP8 version; Unsloth has released a 4-bit version with NVFP4 quantization.76% similarUnverifiedCommon quantization formats include FP32, FP16/BF16, INT8, INT4, and even INT3/INT2.72% 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 BF16.71% similarUnverifiedThe solution uses BF16 precision uniformly for the KV cache, and MXFP4 quantization for the model weights.71% similarUnverifiedThe Qwen3.8-27B model is available in multiple formats including Safetensors, GGUF, NVFP4, NVFP4 GGUF variants, and GPTQ-Int4 quantized versions.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1049448API
curl https://kongchang.com/api/v1/knowledge/claims/1049448MCP
get_claim(id=1049448)