Tag
KittenML 开源发布了 Kitten-TTS 2，这是全球首个 1.58-bit（三值化权重）TTS 模型，1.7B 参数，可在纯 CPU 上实时推理，支持少于 5 秒参考音频的零样本语音克隆、20+ 语言多语种支持以及情感与强调控制。
This preprint argues that learned embedding tables dominate small language models' size and shows a frozen 16 MB vocabulary of 512-bit codes can match or beat trained tables on benchmarks, enabling a 46 MB BitNet ternary model with 5-6x faster CPU inference at a seventh of the memory.
PrismML 推出基于 Qwen3.5 的 Ternary Bonsai 2 系列新模型 2B、4B 和 9B，采用 {−1, 0, +1} 三值权重表示，内存占用比全精度模型小 9 倍以上，并在吞吐与能效上最高提升 5 倍和 3 倍，面向手机、电脑和可穿戴设备上的本地智能体场景。
作者将 Ternary Bonsai 2 27B 量化模型迁移到 Apple Neural Engine，在 16GB M6 Mac mini 上实现约 52 tokens/s 的代码生成速度，整机功耗仅约 12W，显著优于 GPU 方案。项目基于 ANEMLL Forge 提供转换脚本、运行时补丁和 OpenAI 兼容接口，并已适配 Grok Build。
Scion-35B-A3B packs a 35B-A3B MoE into a single 11.34 GB GGUF at 2.61 bpw using ternary expert banks plus trained rank-512 corrections, achieving the best perplexity in the 2-bit class with Q4-class task retention on llama.cpp.
Ternary-Bonsai-2-27B is a 27B parameter AI model compressed into 2-bit ternary format, running on llama.cpp with CUDA and Metal support, facilitating lightweight on-device AI deployment.
The article reveals that Bonsai's headlines are cherry-picked, showing Ternary Bonsai 2 27B retains about 75% of full-precision performance compared to Qwen models on Terminal-Bench and SWE-bench benchmarks.
Ternary Bonsai 2 is a 27B parameter model derived from Qwen3.8-27B that uses ternary weights to achieve a size under 6GB while retaining 98.2% of its intelligence, enabling it to run in-browser on WebGPU.
Prism ML released a ternary weight 27B-class AI model optimized for on-device use on Apple laptops, retaining 98.2% of full-precision intelligence with an 8.60 GB footprint and ~47 tok/s performance.
This research report evaluates post-training ternarization of the Qwen3-4B model, achieving a 1.641-bit effective weight representation with substantial storage compression, while noting a performance trade-off and unresolved deployment acceleration issues.
QTEA is a sub-2-bit post-training quantization framework for large language models that achieves ternary weight compression with improved accuracy and hardware efficiency through sparse residual salient weights and column-wise optimization.
This research paper analyzes capability-stratified degradation in ternary quantized language models, showing that while factual knowledge deteriorates significantly, commonsense reasoning and downstream task adaptability are retained, making the models viable for efficient edge deployment.
Introduces Tritium, an open-source Rust/CUDA engine for ternary (1.58-bit) quantization, serving, and training of LLMs on consumer GPUs. It claims faster inference than llama.cpp for ternary models and introduces a new quantization method called SALT.
Prism ML released Ternary-Bonsai-27B, a ternary-quantized version of Qwen3.6-27B that retains 95% of FP16 intelligence at a ~7.2 GB footprint, enabling full 27B-class reasoning on laptops and single GPUs with speeds up to 26 tok/s on Apple M5 Pro.
Clark Labs released Clark Air Sana 1.6B, a ternary-quantized version of the Sana 1.6B text-to-image transformer that is 8.6× smaller than FP16 while maintaining near-FP16 quality, enabling efficient deployment.
CAT-Q introduces a post-training ternary quantization method for LLMs that uses learnable modulation and softened ternarization, achieving superior performance over BitNet 1.58-bit while using only 512 calibration samples and scaling to 235B parameters.
OpenBMB releases BitCPM4-CANN, a collection of natively trained 1.58-bit ternary quantized LLMs (0.5B to 8B) optimized for Ascend NPUs via CANN, achieving 6× memory reduction at inference and minimal training overhead.
This paper introduces Tequila, a trapping-free quantization method for Large Language Models that improves ternary quantization accuracy and inference speed by repurposing deadzone-trapped weights as dynamic biases.