BitNet b1.58 2B4T

BitNet b1.58 2B4T is Microsoft Research's proof that a language model can be trained natively in ternary, with every weight restricted to -1, 0 or 1, and still hang with full-precision models of the same size. The checkpoint packs 2 billion parameters trained on 4 trillion tokens, and because the non-embedding memory footprint is around 0.4 GB, it runs on ordinary CPUs, including Apple Silicon, through the bitnet.cpp inference framework. Integer-only arithmetic replaces the usual float matrix math, which is where the energy and memory savings come from.

Microsoft's reported numbers have it beating Llama 3.2 1B, Gemma 3 1B and Qwen2.5 1.5B on GSM8K and PIQA, with up to 2x faster inference than comparable models, though absolute benchmark figures were not published in the sources checked. The practical catch is the toolchain: you need bitnet.cpp builds, GPU support is limited, and the model only really shines on supported CPU targets. Weights and code are MIT licensed, which makes it one of the most permissively licensed frontier efficiency experiments around.

• Provider-reported: outperforms Llama 3.2 1B, Gemma 3 1B and Qwen2.5 1.5B on GSM8K and PIQA (relative claim; no absolute numbers published in sources checked)
• Up to ~2x inference speedup vs comparable models with a fraction of the memory (provider claim, bitnet.cpp)
• No published independent benchmark headlines found