Unverified50% confidenceFactExact time
In FP8 mixed-precision training, V3 uses FP8 computation for most matrix multiplications and retains higher precision for precision-sensitive steps, theoretically doubling compute throughput and halving memory usage compared with BF16.
1
Sources
50%
Confidence
Long-term
Relevance
10/23/2026
First Seen
Sources
Understanding DeepSeek in 3 Minutes: The Rise from a Quant Hedge Fund to an Open-Source AI Leader
bilibiliGerrardYX10/23/2026
Related Entities
Related Claims
UnverifiedGradient checkpointing and mixed-precision training (FP16/BF16) are common memory-reduction measures.77% similarUnverifiedMixed-precision quantization applies aggressive low precision such as INT4 to backbone networks that have large parameter counts but are not precision-sensitive, while retaining INT8 or FP16 for precision-sensitive components such as text encoders and QKV projections.74% similarVerifiedFP8 has two variants: E4M3 (4-bit exponent + 3-bit mantissa), which has higher precision and is suited for forward inference, and E5M2 (5-bit exponent + 2-bit mantissa), which has a larger dynamic range and is suited for gradient computation.74% similarUnverifiedCompared to FP16 original precision, FP8 quantization typically causes no more than 1–2% performance degradation on mainstream benchmarks.73% similarVerifiedCompared to FP16, BF16 has the same 8-bit exponent width as FP32 and therefore a larger dynamic range, making it the mainstream precision choice for current large model training.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1090505API
curl https://kongchang.com/api/v1/knowledge/claims/1090505MCP
get_claim(id=1090505)