Unverified50% confidenceFactExact time
The Tiny version model's performance levels off after about 4K hours of data, while the Small version keeps improving.
1
Sources
50%
Confidence
Long-term
Relevance
11/9/2026
First Seen
Sources
Related Entities
Related Claims
Unverified48GB of memory can fit a 4-bit quantized 32B model, but the inference speed may slow to fewer than 1 token per second, making it nearly unusable for interactive purposes65% similarUnverifiedWhen using two threads on a Raspberry Pi 5, model generation speed was 18% to 38% faster than with four threads, because generation is constrained by memory bandwidth rather than CPU compute.65% similarUnverifiedAt 4-bit precision and above, quantization has limited impact on output quality for most application scenarios.65% similarUnverified微虚拟机的内存占用仅约5MB,启动延迟极低63% similarUnverified缺少独立 GPU 的设备运行 2B 参数模型时生成速度可能低于每秒 5 个 token63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1152666API
curl https://kongchang.com/api/v1/knowledge/claims/1152666MCP
get_claim(id=1152666)