Model Catalog

Smallest available chat model for on-device and extreme-latency scenarios.

Ultra-low latency chat for simple tasks. Ideal for autocomplete, short summaries, and high-volume pipelines. 2-4B edge model.

Ultra-lightweight model for high-throughput, cost-sensitive text generation.

Upgraded compact model with strong multilingual support.

Ultra-fast compact chat. Optimised for low-latency tasks.

Compact open-weight chat with automatic multi-tier failover.

Lightweight instruct chat with automatic multi-tier failover.

Highest-throughput free chat tier for high-frequency lightweight tasks. Automatic multi-tier failover.

Balanced MoE chat with automatic multi-tier failover.

Efficient general chat with automatic multi-tier failover.

Indic-languages-optimized chat with automatic multi-tier failover.

Japanese-optimized chat with automatic multi-tier failover.

Multilingual frontier chat with automatic multi-tier failover.

Long-context agentic chat with automatic multi-tier failover.

Classic sparse-MoE chat with automatic multi-tier failover.

Quality between turbo and pro tiers.

Flagship MoE chat with automatic multi-tier failover.

Frontier-grade 550B MoE general chat with automatic multi-tier failover.

Broad-capability open-weight chat. Free tier.

Mixture-of-Experts model optimized for instruction-following, function calling, and agentic tool use. Best speed-to-capability ratio.

Advanced conversational AI with frontier-class reasoning. 128K context window with 500+ tokens/second throughput.

Enhanced 70B model with stronger instruction-following, code comprehension, and multi-language support. Ideal for agentic workflows.

High-capacity 120B chat model. Strong reasoning and instruction following.

Highest-quality free chat tier. Best for complex reasoning and long-context tasks.

Fast, multimodal chat with 1M-token context. Free tier with automatic multi-tier failover.

Mixture-of-Experts smart routing. Automatically selects the best model from the pool based on task type, load, and quality requirements. Ideal for unpredictable workloads.

Frontier-grade 675B MoE model for the most demanding reasoning, analysis, and generation tasks. Ideal for enterprise workloads requiring maximum accuracy.

Fast, low-latency MoE chat with automatic multi-tier failover.

High-quality MoE chat with automatic multi-tier failover.

Active speaker detection — identifies which person is speaking in broadcast and video conference streams.

Sparse end-to-end autonomous-vehicle perception and planning model.

Enhanced 70B code model fine-tuned for code generation, summarization, and multi-language tasks including Python, Java, Go, and Rust.

Instant code completion and simple generation. Optimized for IDE integrations and high-throughput code suggestion pipelines.

Strong general coder, faster than the flagship tier.

Expert code generation model optimized for software development. Generate, debug, and refactor code in 90+ programming languages. 128K context window.

World-class 480B agentic coding model with 256K context. Tops leaderboards for software engineering tasks, browser use, and complex multi-file refactors.

MisarCoder free MoE pipeline tuned for code. Routes across multiple expert tiers based on task size and latency target.

Text embeddings with high retrieval accuracy for RAG, semantic search, and classification. Supports long documents.

State-of-the-art multilingual text embeddings with 1024 dimensions. Supports 100+ languages for semantic search and retrieval.

MTEB top-ranked embedding model. 3072-dim Matryoshka Representation Learning (MRL). Free tier with automatic failover.

Specialized code embedding model optimized for code search and retrieval. Supports text, code, and hybrid queries across 20+ programming languages.

Content-safety classifier. Dedicated safety model; no chat fallback.

Topic-control classifier. Flags off-topic or policy-violating conversations.

Content safety classifier. Returns safe/unsafe labels. Applied automatically on open-source model routes.

Advanced content safety model detecting harmful content across 14 categories. Fast, accurate, and production-ready for content moderation.

Multi-modal content safety classifier covering text and images. Identifies unsafe inputs and outputs across 13+ safety categories.

Detect and classify Personally Identifiable Information (PII) in text. Identifies names, emails, phone numbers, addresses, and 20+ PII categories.

Context-aware content safety model with reasoning capabilities. Applies domain-specific policies and explains safety decisions.

Compact reasoning with automatic multi-tier failover.

Fast reasoning with automatic multi-tier failover.

Multimodal reasoning with automatic multi-tier failover.

Rapid reasoning with lower latency than v3.

Fast mid-tier reasoning model.

Deep multi-step reasoning with automatic multi-tier failover.

Deep reasoning model with 230B parameters excelling in coding, complex reasoning, and office productivity tasks.

Fast reasoning model built on a 200B sparse MoE architecture. Excels at multi-step math, logic, and scientific problem-solving with chain-of-thought.

High-precision document reranking for improved search relevance. Supports up to 1000 documents per request with multilingual support.

GPU-accelerated passage reranker providing probability scores for QA relevance. Ideal for two-stage retrieval pipelines.

End-to-end protein 3D structure prediction from amino acid sequences. Drug discovery and structural biology.

Protein language model generating high-quality embeddings from amino acid sequences. Ideal for drug discovery, protein function prediction, and bioinformatics pipelines.

Ultra-fast speech-to-text transcription supporting 100+ languages. Industry-leading accuracy with 10× realtime processing speed.

Neural machine translation across 12 language pairs. Supports few-shot example prompts for domain-specific terminology control.

High-quality neural text-to-speech with 300+ natural-sounding voices across 100+ languages. Speed and pitch control included.

Zero-shot voice cloning TTS — generate expressive speech in any voice from a short audio sample. No fine-tuning required.

Physical-world simulation model for generating realistic synthetic video data.

Physics-aware world-state video generation from text prompts and spatial control inputs. Designed for physical AI development and simulation.

Efficient vision-language model with automatic multi-tier failover.

Compact vision-language model with automatic multi-tier failover.

Multilingual multimodal model with 128 experts and 17B active parameters. Processes images and text for detailed visual analysis, OCR, and multimodal Q&A.

Premium multimodal model accepting image and audio inputs. Best for complex image reasoning, document analysis, and audio-visual tasks.

Advanced multimodal vision model with image understanding, OCR, visual Q&A, and document analysis capabilities. 128K context window.

High-quality multimodal understanding.

Real-time voice-optimised conversational model for voice assistants and telephony.

Enhance speech recorded in noisy or reverberant environments to studio quality. Ideal for podcast cleanup, call center audio, and voice applications.