Managed inference for any model.
Serve catalog models or your own weights through one OpenAI-compatible API, on GPUs inside your security boundary.
GraphNInference
Built-in, imported, and custom-deployed models for your agents
▶ Spotlight
Qwen3.8 27B
Dense 27B vision-language model for coding, agentic work, and image/video understanding
Qwen3 235B Instruct
qwen3-235b
Largest open-weight MoE model (22B active) for complex instruction following, agents, and long-context tasks
Qwen3.5 122B
qwen3.5-122b-a10b-fp8
Unified vision-language MoE with 10B active parameters for reasoning over images and video
gpt-oss 120B
gpt-oss-120b
Open-weight Apache-2.0 MoE reasoning model with adjustable reasoning depth
Nemotron 3 Super
nemotron-3-super
NVIDIA hybrid Mamba-2 and MoE model for agents, tool use, and long-context reasoning
Qwen3 Coder
qwen3-coder
Code-specialized MoE model for agentic software engineering and repo-scale context
Qwen3 80B
qwen3-80b
Sparse MoE; the fast default for chat and agent loops
Gemma 4
gemma-4
Google's open-weight Gemma 4 for multilingual chat and long context
Qwen3 VL
qwen3-vl
Vision-language model for image and video understanding, grounding, and OCR
Qwen2.5 VL 7B
qwen2.5-vl-7b-instruct
Compact model for image, document, chart, and video understanding
Model launches
Aug 14, 2026Qwen3.8 27B (image+video)Coding, long-horizon agents and image/video in one dense modelApr 22, 2026Gemma 4 31B InstructA non-Qwen alternative with unusually wide language coverageApr 20, 2026Qwen3-VL Embeddings (4096 dim, multimodal)Puts text, images, screenshots and video in one vector spaceApr 20, 2026Qwen3-VL Reranker (multimodal)Scores relevance across modalities, text against images and backApr 20, 2026Gemma 4 E4B (text+image+video+audio)The catalogue's audio-in model - speech and images without a separate ASR stepApr 15, 2026Qwen3 ASR 1.7B (52 languages)Speech to text for voice agents, streaming or batchMar 11, 2026Qwen3 235B InstructHighest-quality answers when accuracy matters more than costMar 11, 2026Qwen3 CoderCoding agents and IDE workflows at repo-scale context
The built-in catalog
Qwen3 80B Instruct (131K context)High-performance reasoning model for complex tasks with extended contextInput $0.20/1M · Output $0.60/1M
qwen3-80b131K contextQwen3 235B InstructLargest open-weight MoE model (22B active) for complex instruction following, agents, and long-context tasksInput $0.22/1M · Output $0.88/1Mqwen3-235b262K contextQwen3 CoderCode-specialized MoE model for agentic software engineering, repo-scale context, and multi-turn coding workflowsInput $0.90/1M · Output $0.90/1Mqwen3-coder131K contextNemotron 3 Super 120BNVIDIA hybrid Mamba-2 / MoE / attention model (12B active) for agentic workflows and long-context reasoningInput $0.90/1M · Output $0.90/1Mnemotron-3-super262K contextgpt-oss 120BOpen-weight Apache-2.0 MoE reasoning model with adjustable low/medium/high reasoning effort and full chain-of-thought via the harmony parserInput $0.15/1M · Output $0.60/1Mgpt-oss-120b131K contextQwen2.5 VL 7B (image+video)Compact vision-language model for image, document, chart, and video understandingInput $0.05/1M · Output $0.05/1Mqwen2.5-vl-7b-instruct128K contextQwen3.5 9B Vision (image+video)General-purpose 9B vision-language model for image analysis and multimodal tasksInput $0.10/1M · Output $0.15/1Mqwen3-vl262K contextQwen3.8 27B (image+video)Dense 27B vision-language model for coding, agentic work, and image/video understandingInput $0.45/1M · Output $3.20/1Mqwen3.8-27b262K contextQwen3.5 122B Vision (image+video)Unified vision-language MoE (10B active) for high-quality multimodal reasoning over images and videoInput $0.29/1M · Output $2.40/1Mqwen3.5-122b-a10b-fp8262K contextDeepSeek OCRDeepSeek OCR model for document-to-markdown conversionPricing unavailabledeepseek-ocrOCRHunyuanOCRHunyuan OCR model for document-to-markdown conversionPricing unavailablehunyuan-ocrOCRQwen3 ASR 1.7B (52 languages)SOTA open-source ASR supporting 52 languages including JapanesePricing unavailableqwen3-asrSpeech-to-textQwen3 TTS (10 languages, 9 speakers)SOTA open-source TTS with custom voice, voice cloning, and voice design across 10 languages including JapanesePricing unavailableqwen3-ttsText-to-speechBGE-M3 Embeddings (1024 dim)Multilingual embeddings for semantic searchPricing unavailablebge-m3EmbeddingQwen3-VL Embeddings (4096 dim, multimodal)Multimodal embeddings (text + image) for knowledgebase image/text searchPricing unavailableqwen3-vl-embeddingEmbeddingBGE Reranker v2Cross-encoder reranker for search result refinementPricing unavailablebge-rerankerRerankerQwen3-VL Reranker (multimodal)Cross-modal reranker for text-to-image and image-to-text relevance scoringPricing unavailableqwen3-vl-rerankerRerankerGemma 4 31B InstructOptimized Gemma 4 31B with FP8 weights, MTP speculative decoding, and 131K contextInput $0.39/1M · Output $0.97/1Mgemma-4131K contextGemma 4 E4B (text+image+video+audio)Edge-optimized Gemma 4 multimodal model with native audio (conformer), image, and video understandingInput $0.15/1M · Output $0.60/1Mgemma-4-e4b66K contextEvery model has its own page with strengths, caveats, and a runnable request. The live list is in the app under Models, or via graphn model list-chat.
Bring your own model
Give the SDK a Hugging Face repo or an S3 path (LoRA adapters work too) and GraphN deploys the weights on managed GPUs, behind the same API as the catalog. There is no serving stack to build.
If a model already runs elsewhere (vLLM, Ollama, OpenAI, Gemini, OpenRouter), register the endpoint and agents use it like any other model.
Custom model imports →S3 weight setup →GraphNInference
Built-in, imported, and custom-deployed models for your agents
my-llama
custom weights · huggingface
Imported from Qwen/Qwen3-0.6B; serving on managed GPUs behind the catalog API
$ graphn model deployHugging Face or S3 weights, LoRA adapters included.import graphn
Call it from anywhere
Chat completions use OpenAI's request shape, so existing tooling works unchanged for catalog and custom models. The Python SDK adds model aliases:
from graphn import chat
reply = await chat.complete(
"Explain async/await in one sentence.",
model="qwen3-235b",
temperature=0.2,
)