GPT-OSS 120B

117B-parameter MoE with 5.1B active per token. Configurable reasoning effort, full chain-of-thought, and native tool use. Apache 2.0.

DeepSeek V3.1 FP4

671B-parameter MoE hybrid reasoning model with 37B active per token. Hybrid thinking mode, Multi-Latent Attention, 160 routed experts.

DeepSeek V3.2 FP4

685B-parameter MoE with 256 routed experts, 8 active per token (37B active). DeepSeek Sparse Attention, integrated reasoning and tool-use.

Qwen 3 Next 80B A3B Thinking FP8

80B total with only 3B active per token using high-sparsity MoE. Hybrid Attention (Gated DeltaNet + Gated Attention), Multi-Token Prediction.

Qwen3 235B A22B Thinking 2507 FP8

235B total with 128 experts, 8 active per token (22B active). State-of-the-art reasoning with extended thinking mode. 256K context.

Qwen3 235B VL Instruct FP8

235B MoE model optimized for general-purpose text generation without thinking blocks. Enhanced instruction following, 256K context.

Qwen3 Coder 480B FP4

480B-parameter MoE with 160 experts, 8 active per token (35B active). SOTA agentic coding, repo-scale understanding, Fill-in-the-Middle.

Qwen3 VL 30B Thinking FP4

30B-parameter dense vision-language model with dedicated thinking capability. Vision-language reasoning with chain-of-thought.

Qwen3 VL 30B Instruct FP4

30B-parameter dense vision-language model optimized for direct instruction following. Broad visual tasks, OCR, and image understanding.

Qwen3 30B A3B FP8

Compact MoE with 30B total and 3B active per token. Dual mode (thinking/non-thinking), 100+ languages.

Qwen3 8B FP8

8B dense model optimized for low-latency responses and lightweight deployment. Dual mode, 100+ languages.

Qwen3.5 397B A17B FP4

Native multimodal foundation model with hybrid Gated Delta Networks and sparse MoE (512 experts). 397B total / 17B active. 201 languages.

MiniMax M2.5 FP8

230B-parameter MoE with 10B active per token. Lightning Attention, 80.2% SWE-Bench Verified, trained on 200K+ real-world environments.

Kimi K2.5 FP4

1T-parameter native multimodal agentic model with 32B active per token. Agent Swarm (100 sub-agents, 1,500 tool calls). Modified MIT.

Llama 3.1 8B

Lightweight 8B dense transformer with Grouped Query Attention. 128K context, 8 languages. Llama Community License.

Llama 3.3 70B FP8

Balanced 70B dense model with 405B-level performance. Grouped Query Attention, 128K context, 8 languages.

Llama 4 Scout FP8

Efficient MoE with 16 experts, 1 active per token (109B total / 17B active). 10M token context, native multimodal, early fusion.

Llama 4 Maverick FP8

Large-scale MoE with 128 routed experts (400B total / 17B active). 1M token context, native multimodal.

Eigen Image

Fast text-to-image generation powered by Eigen AI. Transform text prompts into high-quality images.

Qwen Image Edit 2511

Advanced image editing with auto pipeline selection - supports lightning and standard modes

wan2.2 I2V A14B Turbo

Image to video generation with prompt-guided motion

Whisper V3 Turbo

Convert speech to text using advanced Whisper AI. Record audio or upload files for transcription.

Higgs Audio V2.5

Transform text into natural-sounding speech with multiple voice options and styles.

ChatterBox Voice Twin

Create expressive text-to-speech with voice cloning. Upload a reference sample or use the default Eigen AI voice.

Qwen3 TTS

Advanced text-to-speech with multilingual support and voice cloning. Supports named speakers and custom voice cloning via uploaded reference audio.