Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.
Analysis10
github.com2026-09-28authorityHN front pageabout the bitnet familyHN 150 pts · 31 comments632 words
A distributed pipeline inference engine runs a sliced 0.5B LLM across a cluster of 7 ESP32s3 nodes using 1.58-bit (BitNet) ternary quantization. The master node manages INT4 embeddings (~14MB in Flash) and tokenization while compute nodes execute attention and MLP layers via high-speed SPI daisy-chain communication. This project is licensed under the MIT License.written, reviewed, cited
- The project uses Quantization-Aware Training fine-tuning to adapt the model for 1.58-bit ternary quantization.
- Vocabulary pruning scales the token set down to 32K tokens to fit the embedded hardware constraints.
- Node firmware relies on assembly-optimized MAC operations for efficient 1.58-bit linear layer execution.
github.com2026-08-05authorityHN 1 pts · 1 comments986 words
geistlib is a dependency-free C23 inference engine that runs Microsoft BitNet 2B locally on a 4 GB Raspberry Pi 5. The 1.1 GiB ternary model achieves 15–18 decode tokens/s using ARM SDOT on the Pi and AVX-512 VNNI on x86 without a GPU. It ships under the Apache License 2.0 with prebuilt binaries for Linux arm64 and x86_64.written, reviewed, cited
- The engine supports Qwen3.5, Qwen3.6, and Qwen3.8 dense hybrids including Ternary-Bonsai-2-27B alongside standard families.
- macOS builds require gcc version 14 or Apple clang version 16, plus Homebrew libomp for multi-threading support.
- The GPU Metal backend achieves 1.41 times faster decoding than llama.cpp on an M1 Max when running Qwen3.8-27B.
github.com2024-10-18authorityHN front pageabout the bitnet familyHN 370 pts · 167 comments1,883 words
bitnet.cpp is the official inference framework for 1-bit LLMs that achieves speedups of 1.37x to 5.07x on ARM CPUs and 2.37x to 6.17x on x86 CPUs. The framework reduces energy consumption by 55.4% to 70.0% on ARM and 71.9% to 82.2% on x86, enabling a 100B BitNet b1.58 model to run on a single CPU at 5-7 tokens per second.written, reviewed, cited
- The framework supports lossless inference using 2 bits per weight for embedding models.
- VibeASR.cpp achieves real-time multilingual ASR inference with an RTF less than 1 on x86 and ARM platforms.
- The official 2B parameter model was trained on 4 trillion tokens.
arxiv.org2025-04-17authorityHN front pageHN 111 pts · 30 comments322 words
BitNet b1.58 2B4T is the first open-source, native 1-bit Large Language Model at the 2-billion parameter scale. Trained on a corpus of 4 trillion tokens, the model achieves performance on par with leading open-weight, full-precision LLMs of similar size while offering significant advantages in computational efficiency. The model weights are released via Hugging Face along with open-source inference implementations for both GPU and CPU architectures.written, reviewed, cited
github.com2026-08-02authorityabout the bitnet familyHN 1 pts · 0 comments888 words
Tritium is an Apache-2.0 library for converting dense neural networks into compact additive-ternary models using {-1, 0, +1} weights. The repository currently uses the 1.1.0-rc.2 candidate version, with flagship targets like Qwen/Qwen3.6-27B pending quality gates. Build defaults to CPU-only via cargo build --locked, while CUDA adapters require feature flags.written, reviewed, cited
- Trainable whole-model ONNX generation is intentionally deferred until version 1.3.
- Browser support for Tritium awaits physical receipts from Chrome, Firefox, and Safari.
- The v1.0 codebase contains recorded evidence for BitNet b1.58 2B4T on CPU and CUDA.
huggingface.co2024-03-23authorityabout the bitnet familyHN 11 pts · 1 comments536 words
github.com2026-03-11authorityabout the bitnet familyHN 2 pts · 1 comments830 words
arxiv.org2023-10-18authorityabout the bitnet familyHN 6 pts · 0 comments304 words
arxiv.org2024-11-09authorityabout the bitnet familyHN 3 pts · 0 comments315 words
github.com2024-06-27authorityabout the bitnet familyHN 3 pts · 0 comments473 words
