Training models with ternary quantized weights using PyTorch
-
Updated
Jun 12, 2019 - Python
#
ternary-quantization
Here are 22 public repositories matching this topic...
n0x CLI is my go-to Claude Code alternative when I’m on a budget.
nodejs cli open-source npm typescript terminal ai developer-tools code-generation quantization bonsai low-resource ternary-quantization llm llama-cpp local-llm ollama react-agent offline-ai coding-agent
-
Updated
Jul 4, 2026 - TypeScript
Custom Apple Metal kernels for Hy4-preview (770B ternary MoE) in llama.cpp — makes AngelSlim's STQ1_0 GGUF runnable on Apple Silicon
-
Updated
Sep 2, 2026 - Metal
Pre-quantized ternary models for VLMs, multimodal, and audio — the models GGUF can't touch
quantization whisper gemma vlm model-compression multimodal edge-ai ternary-quantization huggingface apple-silicon vision-language-model qwen
-
Updated
Apr 27, 2026 - Python
笔记本 RTX 5070 Ti Laptop 12G（12,227 MiB / 140 W TGP）· Bonsai-2-27B 三元量化 · MTP vs DFlash2 同上下文 A/B。Laptop GPU only — NOT the desktop 5070 Ti (16 GB / ~300 W); numbers are not comparable across the two.
benchmark cuda mtp bonsai ternary-quantization llm-inference speculative-decoding sm120 rtx-5070-ti dflash2 ninfer
-
Updated
Sep 28, 2026 - PowerShell
Deploy Bonsai-2-27B (ternary 27B, 5.95GB) on a single 8GB consumer GPU via an AI agent. 用 AI 部署 Bonsai-2-27B：一条提示词让 AI 装驱动、编 CUDA、下模型并接入 Claude Code。实测 51 tok/s @4k / 39 tok/s @100k，含按机器调参指南与完整测评报告。
deployment cuda nvidia quantization ternary bonsai ternary-quantization ai-agent llm llama-cpp local-llm llm-inference local-ai gguf claude-code qwen3 consumer-gpu rtx-3070 bonsai-2 bonsai-2-27b
-
Updated
Oct 6, 2026 - Shell
1.58-bit ternary Mamba LLM for Indian languages. Weights are {-1,0,+1} — inference uses only add/sub. 3B model fits in 750MB, runs 20+ tok/s on mobile.
-
Updated
Apr 10, 2026 - Python
QTEA: Ternary LLMs with Sparse Residual Salient Weight and By-Column Optimization
cuda quantization ternary model-compression ternary-quantization post-training-quantization low-bit-quantization llm gptq llama3 llm-quantization qwen3 emnlp-2026 sub-2-bit
-
Updated
Sep 28, 2026 - Python
First open-source qutrit (d=3) variational quantum classifier for tabular data. Lossless 2³→3² binary-to-ternary compression, CSUM ring entanglement, and SHAP explainability. 72 parameters. Built with PennyLane + PyTorch.
pytorch model-compression quantum-machine-learning qiskit explainable-ai shap ternary-quantization pennylane variational-quantum-classifier vqc ternary-computing qutrit ternary-llm
-
Updated
Oct 3, 2026 - Dart
Selective ternary compression of Qwen3.6 MoE expert projections on Apple Silicon, with matched-control functional, routing, and perplexity evaluation.
machine-learning quantization mlx model-compression mixture-of-experts ternary-quantization apple-silicon large-language-models local-llm qwen
-
Updated
Jul 31, 2026 - TypeScript
Colab-friendly BitNet distillation engine: collect KD traces from a teacher, train a ternary Mini-BitNet, and dry-run 7B memory. Multi-provider + Drive/S3
transformers pytorch parquet knowledge-distillation model-compression bitnet colab-notebook mixed-precision ternary-quantization large-language-model efficient-llm efficient-llm-inference activation-quantization kd-traces
-
Updated
Sep 5, 2025 - Python
Anonymous paper release for TRACE-Q, a topology-aware ternary quantization framework for heterogeneous language models.
-
Updated
Jul 30, 2026
Research implementation of activation-aware ternary and mixed-bit quantization for Qwen3.8-27B, targeting 7-9 GB text inference on 16 GB GPUs.
transformers pytorch quantization model-compression mixed-precision ternary-quantization large-language-models llama-cpp gguf qwen3
-
Updated
Aug 17, 2026 - Python
Run PrismML Ternary Bonsai 2 27B (PTQ1_0 packed-ternary GGUF) natively on vLLM — tensor-core fused Triton kernel, 5.9GB resident, 890 tok/s @ bs=32, 2.5x official fork with matched slots
-
Updated
Oct 1, 2026 - Python
Pure-PyTorch inference for moondream/parakeet-redux (ternary FastConformer-TDT ASR): no moondream package / Photon runtime. Verified identical to the official runtime (0.0% WER on 6/6 test files).
-
Updated
Oct 6, 2026 - Python
Independent forensics of Bonsai 2 27B ternary quantization: format and basis recovery, the trained-weights residual, and the public PTQ calibration artifact.
reproducibility quantization bonsai model-compression mixture-of-experts ternary-quantization llm gguf
-
Updated
Sep 27, 2026 - Python
Native low-bit (NLT) diffusion models — the weights live in the quantized space from step 0 instead of being compressed after fp16 pre-training. Home of the AquariusImage series: Aquarius Terimage (ternary text-to-image, released) and Aquarius Binimage (planned).
research pytorch image-generation gradio text-to-image from-scratch model-compression efficient-inference diffusion-model deep-learning-research ternary-quantization quantization-aware-training low-bit-quantization int4 stable-diffusion generative-ai native-low-bit-training
-
Updated
Oct 4, 2026 - Python
PILON (Primitive-Induced Linear Operator Network) explores a compositional weight parameterization for transformer FFN layers. The goal is to replace dense FFN matrices with shared low-rank primitives plus learned composition weights.
research pytorch transformer language-model model-compression bitnet weight-sharing low-rank ternary-quantization quantization-aware-training efficient-deep-learning consumer-gpu
-
Updated
Mar 20, 2026 - Python
ESP32 LLM / MoE: trained on Google Colab GPU, ternary-quantized, exported to C++, and verified on real ESP32 hardware. Fork of Ahmed Barakat's ESP-LLM.
machine-learning microcontroller ai esp32 platformio moe language-model google-colab bitnet arduino-esp32 mixture-of-experts embedded-ai edge-ai ternary-quantization tinyml llm tiny-llm esp32-llm on-device-inference embedded-llm
-
Updated
Sep 26, 2026 - C++
jade_mage · agentprivacy dual-agent harness lane for JadeZaher/ternary-memory-research (fork: main tracks upstream; lane on branch harness/jade_mage; contributions go upstream by PR)
-
Updated
Sep 21, 2026 - Python
Add this topic to your repo
To associate your repository with the ternary-quantization topic, visit your repo's landing page and select "manage topics."