Open weights only

The open catalog.

Every model here ships as open weights — which is what makes it deployable inside your boundary. Closed models don't appear because access isn't deployment; for those, use our cloud on sideren.io. Any model below can land in a private or dedicated deployment.

[33] open models[5] modalities+ bring your own weights

text[01/05]

Reasoners, coders and generalists — the models that run agent loops.

DeepSeek-R1DeepSeekFrontier open reasoning — long chains, tool use, agent loops.

DeepSeek-V3.2DeepSeekGeneral flagship — strong code and instruction following at scale.

Qwen3-235BQwenMoE flagship of the largest open family — hybrid reasoning modes.

Qwen3-CoderQwenAgentic coding specialist — repo-scale context, tool calling.

Llama 4 MaverickMetaNatively multimodal MoE — text and image understanding.

gpt-oss-120bOpenAIOpenAI's open-weights reasoner — efficient at production sizes.

GLM-4.7ZhipuStrong agentic and coding line — fast variants for loops.

Kimi K2MoonshotTrillion-parameter-class MoE — built for agentic execution.

MiniMax-M2MiniMaxCompact MoE tuned for agent + code workflows.

Mistral Small 3.2Mistral AIDense workhorse — low latency on modest hardware.

Gemma 3 27BGoogleSingle-accelerator class — multimodal, 140+ languages.

ERNIE 4.5BaiduOpen multimodal MoE family, ten sizes from 0.3B up.

LongCat-FlashMeituanSpeed-first MoE — dynamic activation for cheap tokens.

embeddings[02/05]

Retrieval backbones for search, RAG and memory.

Qwen3-Embedding-8BQwenMultilingual retrieval flagship — instruction-aware.

BGE-M3BAAIDense + sparse + multi-vector in one model, 100+ languages.

EmbeddingGemmaGoogleOn-device class — small, fast, strong for its size.

nomic-embed-text-v2Nomic AIFully open MoE embedder — weights, data and code.

jina-embeddings-v3Jina AITask-LoRA embedder — long context, 89 languages.

image[03/05]

Generation and editing, photoreal to typographic.

FLUX.2 [dev]Black Forest LabsThe open image flagship — generation and editing in one.

SD 3.5 LargeStability AIOpen generation staple — huge ecosystem of fine-tunes.

Qwen-ImageQwenText rendering inside images that actually holds up.

HunyuanImage 3.0TencentLargest open image MoE — photoreal, bilingual prompts.

CogView4ZhipuOpen generation with reliable Chinese + English text.

video[04/05]

Text- and image-to-video, from clips to production shots.

Wan 2.2AlibabaOpen video flagship — text- and image-to-video.

HunyuanVideo 1.5TencentEfficient open video — strong motion, runs lean.

CogVideoXZhipuProven open text-to-video line with wide tooling.

Step-Video-T2VStepFun30B open text-to-video — long clips, two languages.

voice[05/05]

Speech in, speech out — transcription, synthesis, full duplex.

Whisper large-v3OpenAIThe default open speech-to-text, ~100 languages.

Parakeet TDTNVIDIABlazing transcription — hours of audio in seconds.

CosyVoice 2AlibabaStreaming TTS — 150 ms first packet, voice cloning.

Step-Audio 2StepFunEnd-to-end speech understanding + generation.

MiMo-AudioXiaomiFew-shot audio generalist — speech in, speech out.

ChatterboxResemble AIProduction TTS with emotion control — MIT licensed.

your weights

Bring your own model.

Fine-tunes, adapters, internal checkpoints — anything you hold the weights for lands in the same plane, behind the same endpoint, under the same guarantees. Your weights never leave your boundary either.

Scope a deployment

Catalog reflects notable open-weights releases — availability per deployment is scoped with your team, licenses reviewed together.