Inkling enters the open-weight race

Teams that want to adapt multimodal reasoning models now have another model to test on their own data. Thinking Machines wrote on X that “Inkling reasons efficiently across text, image, and audio modalities” and said it is making the full weights available. The substance is not only the model name; it is the release shape. Inkling is being offered for fine-tuning on Tinker, with a playground for hands-on testing before a training run.

The concrete number to watch is context length. Thinking Machines says Inkling is available on Tinker with 64K and 256K token options. That matters for long documents, multimodal conversations, tool use, and agent workflows where context windows often become the practical limit before model quality does. The company also added cookbook recipes focused on audio capabilities and released tooling for sampling and post-training with tool calls, reasoning content, and multimodal inputs.

Thinking Machines has positioned Tinker as infrastructure for model customization, so this tweet fits a broader pattern: put the model in developers' hands, then make adaptation the product. The linked post describes Inkling as its first open-weights model and says both the original checkpoint and an NVIDIA Blackwell-oriented NVFP4 checkpoint are on Hugging Face. The company also lists deployment and integration paths through TogetherAI, Fireworks, Modal, Databricks, Baseten, vLLM, SGLang, llama.cpp, and transformers.

The next test is whether Inkling turns openness into practical substitution. Benchmarks will matter, but buyers will look harder at fine-tuning cost, audio-heavy workflows, tool-calling reliability, and how the 256K option behaves under real enterprise context. The source tweet and model post give developers enough detail to begin that comparison immediately.

#inkling

Thinking Machines opens Inkling weights for multimodal reasoning