What is Sesame CSM-1B?
CSM-1B is the open Conversational Speech Model from Sesame, the US company behind the Maya and Miles voice assistants, released on 13 March 2025. It uses a Llama backbone and a smaller audio decoder that produces Mimi audio codes at 24 kHz. Despite the name, the Hugging Face page lists about 2 billion parameters including the decoder. It has been supported natively in Hugging Face Transformers since version 4.52.1 in May 2025.
CSM is a base model, not a finished voice: with no context it picks a random speaker, and it sounds best when fed the earlier turns of a conversation as text and audio, which also lets it continue a voice from a sample. It speaks English; other languages are weak. It cannot generate text, so it needs a separate language model. The weights are Apache 2.0 behind a click-through gate. The fine-tuned models behind Sesame's own app remain closed.
Is Sesame CSM-1B worth it?
CSM-1B is worth knowing because it is an open model built around conversational context, and it is still the only weights Sesame has released. Pass it the previous turns of a dialogue, each with text, audio and a speaker ID, and it shapes its delivery to fit, which is what made Sesame's demo voices sound present rather than read aloud.
What you download is not that demo. The card says plainly that the open model has not been fine-tuned on any voice; the assistants in Sesame's October 2026 app run on newer, closed versions. Out of the box you get a random voice per call unless you supply context audio, English only, and no published latency, speed or VRAM figures. Twenty months after release, it has 32 fine-tunes on Hugging Face and about 86,000 downloads in the past month, far below VoxCPM2's 977,000.
The ecosystem is the best part. Native Transformers support means batching, CUDA graphs and fine-tuning with the standard Trainer, and DeepInfra serves it through OpenRouter at $7 per million characters, about half of Boson's or Fish Audio's $15. It does not appear on Artificial Analysis's arena, so there is no independent quality figure.
The catch: safety rests on wording. The card prohibits mimicking real people without explicit consent, fake news and fraudulent calls, and calls the model research and education only, yet the Apache licence places no such limits and the code does not enforce them. The GitHub repository contains a watermarking file, but its behaviour is not documented. The last commit to the repository was in May 2025.
Our take: CSM-1B is a good starting point for a team fine-tuning its own conversational voice in English; anyone wanting a ready voice today will get further with a newer model.
Where it earns its place
A voice-assistant startup with an in-house ML engineer
Fine-tune a branded English voice that reacts to the flow of a conversation
Record the voice talent, fine-tune with Transformers' Trainer, and feed previous turns as context at inference. Apache 2.0 allows shipping the result commercially.
A conversational-AI researcher
Study how dialogue context changes prosody
Generate the same line with and without earlier turns as Segments and compare. The 24 kHz Mimi codec output and speaker IDs make controlled experiments straightforward.
A developer prototyping a phone bot without GPUs
Try CSM's delivery before committing to self-hosting
Call it through OpenRouter's speech endpoint, served by DeepInfra at $7 per million characters, so 10 million characters of tests cost about $70.
Before you commit
- Setup
- Accept the gate on Hugging Face (and access to Llama-3.2-1B for the original repo), then load CsmForConditionalGeneration in Transformers 4.52.1 or later. A CUDA GPU is expected; tested on CUDA 12.4 and 12.6.
- Integrations
- Hugging Face Transformers, pipeline and Trainer; DeepInfra via OpenRouter's /api/v1/audio/speech endpoint returning mp3 or pcm.
- Data and privacy
- Self-hosted, so audio stays local. Through OpenRouter, requests go to DeepInfra and are not available on EU in-region routing. Prohibited uses are stated in the card, not enforced.
- Cost at scale
- Free weights; hosted at $7 per million characters with no OpenRouter markup.
- Limits
- English only; needs context audio for a consistent voice; cannot generate text; no published latency or memory figures.
What’s good
- Uses conversation history to shape delivery
- Apache 2.0 weights
- Native Transformers support with batching and fine-tuning
- Hosted at $7 per million characters via OpenRouter
What’s not
- English only; other languages 'likely won't do well'
- Base model: random voice without context audio
- No published latency, speed or VRAM figures
- No repository commits since May 2025
Best for
Research and product teams that want to fine-tune their own English conversational voice on an open base with a permissive licence.
Sesame CSM-1B alternatives
Sesame CSM-1B: common questions
- Is Sesame CSM-1B worth it?
- CSM-1B is a capable research base for conversational speech, but in late 2026 it is an older, English-only model that needs work before it sounds like Sesame's demo. We rate it 3.3 out of 5.
- How much does Sesame CSM-1B cost?
- Free under Apache 2.0 after accepting the gate. Hosted on DeepInfra through OpenRouter at $7 per million characters.
- What are the downsides of Sesame CSM-1B?
- English only; other languages 'likely won't do well'; base model: random voice without context audio; no published latency, speed or VRAM figures; no repository commits since May 2025.
- Who is Sesame CSM-1B best for?
- Research and product teams that want to fine-tune their own English conversational voice on an open base with a permissive licence.
- How much does Sesame CSM-1B cost for a team or in production?
- Free weights; hosted at $7 per million characters with no OpenRouter markup.
- Is Sesame CSM-1B safe for company data?
- Self-hosted, so audio stays local. Through OpenRouter, requests go to DeepInfra and are not available on EU in-region routing. Prohibited uses are stated in the card, not enforced.
- What are the best Sesame CSM-1B alternatives?
- The closest alternatives we have reviewed are Higgs TTS 3, VoxCPM2, Chatterbox and Maya 2 Native.
Answers checked Nov 7, 2026.