Sesame CSM SesameAILabs/csm

Sesame's 1B Conversational Speech Model that generates speech from text plus optional conversational audio context.

Creator kit + Compare GitHub ↗ Install guide (official) ↗

Sesame CSM at a glance

Checked 15 Oct 2026

CSM-1B (released March 2025) generates speech as Mimi audio codes from text and optional audio context, using a Llama backbone. Code and CSM-1B weights are Apache-2.0, but the README requires a CUDA GPU and access to two gated Hugging Face repos, including meta-llama/Llama-3.2-1B (Llama 3.2 Community License) for the tokenizer. It ships no preset voices, is mainly English, and cannot generate text. Repo last pushed May 2025.

Cite this pageRepoQuarry. “Sesame CSM.” Checked 15 Oct 2026. https://repoquarry.dev/projects/sesame-csm

How much of ElevenLabs does it replace?

1 of 4 jobs. Base model with no fine-tuned voices; README says non-English output likely won't do well.

Speech generation Partial replacement

Voice cloning Not covered

Speech recognition Not covered

Voice agents Not covered

For voice cloning, speech recognition, voice agents, see other ElevenLabs alternatives. You may also need to build an interface, an API server and hosting yourself.

Facts, with sources

Every answer links to where we found it. We read the documentation; we did not install it.

Report a correction

QuestionAnswerSourceChecked

Runs on my machine? Yes, Python code runs the model on your own machine If you want to write your own applications with CSM, the following examples show basic usage. github.com · csm#usage ↗ 1d ago

Code licence Apache-2.0 github.com · LICENSE ↗ 1d ago

Model weights licence CSM-1B weights Apache-2.0 (gated: accept conditions on Hugging Face); the code also loads the tokenizer from meta-llama/Llama-3.2-1B, a manually gated repo under the Llama 3.2 Community License This repository is publicly accessible, but you have to accept the conditions to access its files and content. huggingface.co · csm-1b ↗ 1d ago

Commercial use Apache-2.0 code and CSM weights allow it, but the required Llama 3.2 tokenizer is under the Llama 3.2 Community License (extra terms, e.g. a separate licence above 700M monthly active users), and the README describes the project as for research and educational purposes with listed prohibited uses If, on the Llama 3.2 version release date, the monthly active users of the products or services made available by or for Licensee, or Licensee’s affiliates, is greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta raw.githubusercontent.com · LICENSE ↗ 1d ago

Paid services needed? None required; a free Hugging Face account with access granted to CSM-1B and Llama-3.2-1B is needed You will need access to CSM-1B and Llama-3.2-1B github.com · csm#requirements ↗ 1d ago

Hardware README lists a CUDA-compatible GPU as a requirement; VRAM and RAM amounts are not stated (example code also selects MPS or CPU if no CUDA) A CUDA-compatible GPU github.com · csm#requirements ↗ 1d ago

Operating systems Windows setup notes given (use triton-windows); Linux and macOS are not explicitly stated The triton package cannot be installed in Windows. Instead use pip install triton-windows. github.com · csm#windows-setup ↗ 1d ago

Setup effort Medium; git clone, Python 3.10 venv, pip requirements, CUDA, Hugging Face login for gated models The code has been tested on CUDA 12.4 and 12.6, but it may also work on other versions github.com · csm#setup ↗ 1d ago