Microsoft's BitNet brings 1.58-bit weights to embeddings, 2.28x FP16 throughput on CPU
techNmak · x · 2026-10-30
Microsoft's 41K-star BitNet project has extended 1.58-bit ternary weights to text embedding models, open-sourcing BitNet-embedding-270M (640-dim) and BitNet-embedding-0.6B (1024-dim). Both support sequences up to 32,768 tokens and use ternary weights with 8-bit activations.
In Microsoft's CPU benchmarks, the 600M model reaches up to 2.28x the throughput of its FP16 baseline; with specialized inference kernels this substantially cuts weight storage and memory traffic — attractive for RAG and local AI.
This builds on the same BitNet research behind the 2B language model trained on 4 trillion tokens, where linear-layer weights take only three values: -1, 0, +1.
More from Models
- LLMs are fluent in others' work but unreliable when pressed on details — AlexTensor · 2026-10-31
- 576GB 'go f*ck yourself Sam Altman' mirror spreads as OpenAI docs leak — retr0jirachi · 2026-10-31
- Codex 28-day challenge teases day-28 reset: OpenAI may be saving a new model for the finale — haider1 · 2026-10-31
- GPT-6 ASTRA LAW spotted in OpenAI's API model list, launch looks imminent — imjustnewatai · 2026-10-31
- Grok now accepts Alipay and UnionPay, opening paid access for Chinese users — op7418 · 2026-10-31
- Asking DeepSeek and Hy to predict their own next models yields two very different roadmaps — lucky-adventurer- · 2026-10-31