Inkling is Thinking Machines Lab's open-weight, multimodal foundation model built for fine-tuning by enterprises and developers who want to own and reshape a frontier-grade model instead of renting one. Released on July 15, 2026 by Mira Murati's lab, it is a 975-billion-parameter mixture-of-experts system under the Apache 2.0 license, with weights on Hugging Face.
Core Features
- 975B MoE with 41B active params per token, so it runs cheaper than its size suggests while keeping capacity.
- 1-million-token context window for long documents, codebases, and multi-file reasoning.
- Native multimodal: trained from scratch on 45 trillion tokens spanning text, image, audio, and video, with built-in vision and audio.
- Controllable thinking: dial reasoning effort up or down to trade latency for depth per task.
- Day-0 support in transformers, SGLang, vLLM, and llama.cpp, with MXFP8 and NVFP4 quantized weights available.
Use Cases
- Companies fine-tuning on proprietary data via the Tinker console for a domain-specific model.
- Regulated industries that need a permissive license and on-prem hosting.
- Researchers studying MoE routing and long-context behavior.
- Builders wanting a US-built open base as an alternative to Qwen-class models.
Pricing
Inkling offers a free tier: the weights are free under Apache 2.0. Revenue comes from Tinker, the managed fine-tuning platform; fine-tuned models run at 64K context with a 256K PEFT variant. The hosted API is priced like a frontier model, so most teams self-host the free weights.
Our Take
Best for organizations that need a customizable, license-clean base model and have the data to fine-tune it. The trade-off is that Inkling trails closed models on raw benchmark scores, so it wins on control, not on topping leaderboards. Compare with other AI engine and model options on aifreetool.site.
