DeepSeek-V4-Flash-0731-ROCmFP3

All projects

Apps & Tools

DeepSeek-V4-Flash-0731-ROCmFP3

A high-performance ROCm FP3 quantization of the DeepSeek-V4-Flash model.

bymrciffa

PlayableOpen source

DeepSeek-V4-Flash-0731-ROCmFP3 media is blocked

Allow external media to connect to the provider and play this content.

Description

DeepSeek-V4-Flash-0731-ROCmFP3 is a specialized quantization of the DeepSeek-V4-Flash model optimized for ROCm-compatible hardware. It achieves superior performance on the HellaSwag reasoning benchmark, notably outperforming standard FP8 implementations. The project utilizes deterministic disjoint evaluation shards to validate its accuracy across over 10,000 question samples.

Descriptions, tags, and model credits may be AI-generated or inferred from public sources and can be incomplete or wrong. Learn more.

Details

Why this is interesting: Achieves a technical milestone by delivering an ROCm FP3 quantization that outperforms FP8 on a standard reasoning…
Attributed to: mrciffa@davideciffa
Built with: V4.1 Flash Gemini 3.8 Flash Cyber 3.8 Flash Omni 1.1 Flash Ling 3.0 Flash Fin Gemini 3.7 Flash DeepSeek-V4-Flash-0731 Qwen3.7 Flash 3.6 Flash 3.5 Flash Cyber Step 3.7 Flash Gemini 3.5 Flash V4 2.5 Flash
Links: huggingface.co
View tweet
View GitHub repo
Tags: #deepseek #quantization #rocm #benchmark

Discovered: X
First posted: X · Sep 16, 2026
Listed: Added 2h ago
Build: Strong
Models: Strong