# Tencent/AngelSlim

Model compression toolkit engineered for enhanced usability, comprehensiveness, and efficiency.

Repository: https://github.com/Tencent/AngelSlim
Canonical: https://ross.abutalabs.com/products/angelslim
Homepage: https://angelslim.readthedocs.io
Language: Python
License: NOASSERTION
License Family: other
Topics: llm, llm-compression, quantization, speculative-decoding, diffusion, vlm, hunyuan, deepseek, qwen, fp4, eagle, audio, dflash
Last push: 2026-08-07T13:41:49+00:00

## Health v2 (maintenance only)
Score: 72/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 96, release rhythm 65, longevity 30
- inputs: {"age_days": 426, "days_push": 26, "days_rel": 72, "gap_med": 91, "n_releases_24m": 4}
- flags: no_license
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1547, forks 172 (observed 2026-08-28T04:05:01.611777+00:00)

## What it is
AngelSlim is a Python toolkit from Tencent for compressing large language models and related architectures (VLMs, diffusion, audio models) via quantization, distillation, and speculative decoding. It provides PTQ/QAT pipelines, quantization-aware distillation on Megatron-Core, and acceleration techniques like sparse attention and draft-model-based decoding.

## Use cases
- quantize an LLM to FP8 or FP4 for faster inference
- compress a large model with post-training quantization
- run quantization-aware distillation for Qwen or Hunyuan models
- speed up LLM generation with speculative decoding
- train a draft model for speculative decoding
- accelerate long-context prefill with sparse attention
- distill a full-precision model into a smaller one

## When to choose
- you need to compress or quantize LLMs, VLMs, or diffusion models for deployment
- you want speculative decoding with MTP, EAGLE, or DFly drafters
- you work with Tencent Hunyuan, DeepSeek, or Qwen model families
- you need distributed quantization-aware distillation on Megatron-Core

## When to avoid
- you need compression for small classical ML models rather than large neural networks
- you want a one-click GUI tool with no Python scripting
- your target hardware does not support low-precision formats like FP8/FP4

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, llm-inference, llm-training, deep-learning
- domain: large-language-models, machine-learning, deep-learning, developer-tools
- platform: python
- tags: model-compression, quantization, speculative-decoding, distillation, ptq, qat, sparse-attention, vlm, diffusion-models, gpu, linux

## Member repositories
- Tencent/AngelSlim (main) score 72

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:01.611777+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T04:30:28.033520+00:00, confidence not recorded.
  - readme: https://github.com/Tencent/AngelSlim (fetched 2026-08-28T04:05:01.611777+00:00, sha a2697c2bc263)
  - registry_pypi: https://pypi.org/pypi/angelslim/json (fetched 2026-08-29T11:31:58.735678+00:00, sha e17f9ad9d079)
- Data as of 2026-08-30T08:39:29.467469+00:00.