thinkingmachines/Inkling-2-Small · Hugging Face

Reference

Inkling 2 Small is an open-weight, multimodal AI model that accepts text and images and produces text for chat, coding, agents, and visual tasks. Its large scale and 1-million-token context window make it relevant for developers building advanced AI applications. The model has 594 billion total parameters but activates 17 billion per token through a sparse mixture-of-experts design. It supports English and other languages, image inputs, tool use, retrieval-augmented generation, and multiple coding languages. The recommended NVFP4 version needs at least 500GB of combined GPU memory, such as two NVIDIA B300s or four H200s. An MXFP8 version needs at least 800GB, but the documentation says its evaluation gains are only marginal. Developers can run it with SGLang, vLLM, TokenSpeed, Unsloth, or Hugging Face, and access it through Tinker or listed third-party providers. Training used text, images, video, public sources, third-party data, internally generated data, and synthetic or augmented data. The page reports broad benchmark results across coding, search, business, finance, legal work, visual reasoning, long-context tasks, and factuality. The model is released under the Apache 2.0 license for research, fine-tuning, and integration into third-party products.
  • The model combines a 594-billion-parameter sparse MoE architecture with a 1-million-token context window, but practical deployment requires at least 500GB of aggregated GPU memory.
  • The documentation recommends the NVFP4 checkpoint because the higher-memory MXFP8 version reportedly delivers only marginal evaluation gains.

Build this article’s timeline

Extract dated events, key actors, and open questions using the same timeline analysis available in Signal.

Log in to build

Timeline generation is available to signed-in community members.

Five Ws 🤖 4 AI prefills · 1 still unanswered
Who
Hugging Face, NVIDIA, and Fireworks
🤖 AI prefill
What
Inkling 2 Small is an open-weight, multimodal AI model that accepts text and images and produces text for chat, coding, agents, and visual tasks.
🤖 AI prefill
When When did this happen, or what is the timeframe?
Where
United States
🤖 AI prefill
Why
Centers on Inkling 2 Small, multimodal autoregressive transformer, and sparse Mixture-of-Experts (MoE)
🤖 AI prefill

Login to answer and earn XP for each gap you fill.

Model Capabilities multimodal, text inputs, image inputs, text generation, conversational Deployment Ecosystem Tinker Playground, API, SGLang, vLLM, TokenSpeed Hardware Precision NVFP4, MXFP8, GPU cluster, VRAM, NVIDIA B300 Transformer Architecture 48-layer, decoder-only, Mixture-of-Experts, 384 experts, 8 experts Model Metadata open weights, Apache-2.0, safetensors, downloads, repository
Inkling 2 Small100% multimodal autoregressive transformer98% sparse Mixture-of-Experts (MoE)96% image-text-to-text model94% 48-layer decoder-only transformer92% NVFP4 checkpoint91% MXFP8 checkpoint86% vision transformer88% agentic and tool-use systems84% retrieval-augmented generation83%
Manipulation score look right?

Comments

Login to view and post comments.

🏠Portal 📰Links ❓Q&A 📅Events 💼Jobs