# FasterDecoding/Medusa

Medusa: Simple Framework for Accelerating LLM Generation with Multiple Decoding Heads

Repository: https://github.com/FasterDecoding/Medusa
Canonical: https://ross.abutalabs.com/products/fasterdecoding-medusa
Homepage: https://sites.google.com/view/medusa-llm
Language: Jupyter Notebook
License: Apache-2.0
License Family: permissive
Topics: llm, llm-inference
Last push: 2024-06-25T12:23:04+00:00

## Health v2 (maintenance only)
Score: 18/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 0, release rhythm 8, longevity 77
- inputs: {"age_days": 1088, "days_push": 799, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2770, forks 205 (observed 2026-08-28T04:07:18.888837+00:00)

## What it is
Medusa is a framework that accelerates LLM text generation by adding multiple decoding heads to an existing model, avoiding the need for a separate draft model. It achieves roughly 2-3.6x speedups using tree-based attention and a typical acceptance scheme during decoding.

## Use cases
- speed up llm text generation
- accelerate inference without a draft model
- make local llm hosting faster
- reduce llm generation latency
- add speculative decoding heads to a fine-tuned model
- faster sampling-based generation

## When to choose
- you want faster LLM generation without deploying a separate draft model
- you serve single-batch (batch size 1) inference for local model hosting
- you want parameter-efficient training that leaves the base model untouched
- you need faster non-greedy (sampling) generation

## When to avoid
- you need high-throughput batched serving rather than batch size 1
- you cannot fine-tune or attach new heads to your model
- you need a turnkey inference server rather than a research framework
- your inference stack has no Medusa integration

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, machine-learning, gpu-computing
- domain: large-language-models, machine-learning, deep-learning
- platform: python
- tags: speculative-decoding, decoding-heads, llm-acceleration, tree-attention, inference-optimization, gpu, linux

## Member repositories
- FasterDecoding/Medusa (main) score 18

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:07:18.888837+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 2026-08-30T08:16:40.037230+00:00, confidence not recorded.
  - readme: https://github.com/FasterDecoding/Medusa (fetched 2026-08-28T04:07:18.888837+00:00, sha 7bca66513aee)
  - homepage: https://sites.google.com/view/medusa-llm (fetched 2026-08-29T09:56:11.151359+00:00, sha b568a0587988)
- Data as of 2026-08-30T08:39:29.467469+00:00.