CSM 1B
CSM 1B is a 1‑billion‑parameter conversational speech model from Sesame that turns text (and optionally audio context) into natural‑sounding English speech.
What is CSM 1B?
CSM 1B is a 1‑billion‑parameter conversational speech generation model from Sesame that converts text into English speech. It is mainly used for dialogue-oriented applications such as voice assistants, interactive agents, and chat-style voice interfaces that need realistic, contextual speech output. It is also applied in text-to-speech pipelines for content creation, accessibility tools, and other products that require controllable, high‑fidelity synthetic voices. CSM 1B belongs to Sesame’s CSM (Conversational Speech Model) family, which uses a Llama backbone with a specialized audio decoder that generates RVQ/Mimi audio codes.
Providers
Different companies host the same model. OpenRouter routes your request to one of them based on the routing mode you pick — Balanced (price + speed), Nitro (fastest), or Exacto (highest tool-calling accuracy).
| Provider | Input | Output | Cache read /M | Latency | Throughput | Uptime |
|---|---|---|---|---|---|---|
| Sesame | ~$0.15 | ~$0.15 | — | ~220ms | ~60 tps | ~99.9% |
| OpenRouter | ~$0.18 | ~$0.18 | — | ~250ms | ~55 tps | ~99.9% |
| Replicate | ~$0.20 | ~$0.20 | — | ~280ms | ~45 tps | ~99.5% |
Try this model
Test CSM 1B right here — free to start.
Suggestions for your first prompt
Code snippet
Call the model through the OpenAI-compatible API.
from openai import OpenAI
client = OpenAI(
api_key="YOUR_API_KEY",
base_url="https://inference.example.com/v1"
)
response = client.chat.completions.create(
model="sesame/csm-1b",
messages=[
{
"role": "user",
"content": "Describe this image in one sentence."
}
],
)
print(response.to_json())
{
"model": "sesame/csm-1b",
"messages": [
{
"role": "user",
"content": "Describe this image in one sentence."
}
]
}
5 Core Capabilities
-
Text-to-Speech
Generates high-quality, natural-sounding speech audio directly from text using a Llama-based backbone and specialized audio decoder.
-
Conversational Prosody
Maintains contextual awareness across turns, adjusting tone, pauses, and inflection to match dialogue flow and emotional nuance.
-
Multimodal Inputs
Processes both text and audio inputs, using prior audio context to guide consistent speech patterns and expressive delivery.
-
Streaming Generation
Supports efficient, low-latency speech synthesis suitable for real-time or interactive applications and local deployment scenarios.
-
Multilingual Potential
Can be fine-tuned on additional languages, enabling customized voices and language support beyond the base English-focused model.
6 Most Valuable Use Cases
- Voice Customer Support
- Interactive Voice Assistants
- Audiobook Narration
- Game Character Voices
- Voice Prototyping Tools
- Accessibility Voice Output
Why Build on LLM.API?
One unified API. Every major model. Built-in reliability, cost control, and observability.
-
Unified AI Routing
Dynamically route each request to the best-fit model across providers based on latency, accuracy, or cost—without changing your integration or redeploying code.
One endpoint, every model -
Smart Cost Controls
Optimize spend with per-route price caps, automatic model downgrades, and detailed usage insights so you can keep quality high while staying within budget.
Max quality, min cost -
Automatic Provider Fallbacks
Avoid downtime by failing over to alternative models or providers on errors, rate limits, or outages—configured once, enforced globally in real time.
Resilience by default -
End-to-End Observability
Trace every call across providers with request logs, latency breakdowns, errors, and cost metrics so you can debug faster and tune performance with confidence.
See every token -
Task-Aware Abstractions
Use high-level tasks—chat, tools, RAG, agents—instead of provider-specific APIs, so you can upgrade models or vendors without rewriting your application logic.
Code to tasks, not APIs -
High-Throughput Batch APIs
Run large-scale jobs—evaluations, backfills, fine-tuning prep—through a single batch interface with concurrency, retries, and throttling handled by the platform.
Ship bulk workloads fast
When to Use — When NOT to Use
Use it if...
- You need a compact vision-language model for on-device or edge deployment scenarios.
- You need to classify or tag large volumes of images cost-effectively.
- Your use case involves extracting simple visual attributes or objects from images.
- Your use case involves lightweight multimodal experimentation before scaling to larger Sesame models.
- You need a small model to fine-tune for domain-specific visual recognition.
Avoid if...
- You need state-of-the-art reasoning over complex documents, diagrams, and mixed long-context inputs.
- Your workload requires highly accurate natural language generation beyond short captions or labels.
- You need top-tier performance on intricate multimodal benchmarks or safety-critical decisions.
- Your workload requires robust handling of very high-resolution images without aggressive downscaling.
- You need broad multilingual understanding and generation, not just basic English-centric capabilities.
Frequently Asked Questions
-
What is CSM 1B?
CSM 1B is a 1-billion-parameter language model from Sesame, accessible through LLM.API for lightweight, cost-efficient text generation and understanding.
-
What tasks is CSM 1B best suited for?
CSM 1B is best for lightweight chatbots, autocomplete, short-form content generation, and simple classification or extraction tasks where low cost matters.
-
What is the context window of CSM 1B?
CSM 1B supports a 4K-token context window, making it suitable for short conversations, prompts, and small documents.
-
What modalities does CSM 1B support?
CSM 1B is a text-only model, accepting text prompts and returning text completions without image, audio, or video support.
-
How do I call CSM 1B via LLM.API?
Use the LLM.API completions or chat endpoint with the model parameter set to "Sesame/CSM-1B" and include your usual authorization header.
-
How does CSM 1B compare to larger Sesame models?
CSM 1B is cheaper and faster but generally less capable on complex reasoning, long-context, and nuanced instruction-following than larger Sesame models.
-
What are the typical latency characteristics of CSM 1B on LLM.API?
CSM 1B is optimized for low latency, typically returning first tokens faster than larger models at similar throughput settings.
-
What are the limitations of CSM 1B?
CSM 1B may struggle with long multi-step reasoning, very domain-specific technical tasks, and maintaining consistency over extended dialogs.
-
Does CSM 1B support streaming responses on LLM.API?
Yes, you can enable streaming in LLM.API requests to receive CSM 1B tokens incrementally as they are generated.
-
How is CSM 1B priced on LLM.API?
CSM 1B is priced as a budget-friendly tier on LLM.API, with lower per-token costs than larger Sesame and frontier models.
COMPARE
Competitive Models
-
GPT-5.4 Pro
GPT-5.4 Pro is an OpenAI language model whose specific architecture, capabilities, and release details have not been publicly documented as of now. Any concrete claims about its performance or features beyond official OpenAI announcements would be speculative.
-
GPT-5.4 Nano
GPT-5.4 Nano is an OpenAI model name, but there is no public, reliable information available describing its architecture, capabilities, or intended use. Any additional details would be speculative.
-
Claude Opus 4.6 (Fast)
Claude Opus 4.6 (Fast) is an Anthropic large language model deployment variant that emphasizes reduced latency while retaining strong general-purpose reasoning and generation capabilities. It is designed to provide high-quality answers more quickly than standard Opus configurations.
Get one key to every model
Swap your API key. Keep your code.