Thinky

Thinky is Latent.Space’s informal name for Thinking Machines Lab, the AI company behind Tinker, interaction models, and Inkling. Its work combines infrastructure for customizing models with models built for multimodal reasoning and continuous human collaboration. In July 2026, it released the full weights of Inkling and Inkling-Small, making both available for fine-tuning through Tinker.[1][2][3][5][6][7]

These 2026 releases followed a 2025 period of financing, research, and the launch of Tinker. The founding team was publicly confirmed that July; the subsequent research supplies context for the adaptation methods and numerical infrastructure used in later products.[11][9][10][3][5][6]

On July 15, 2025, WIRED reported that Thinking Machines had raised a $2 billion seed round at a $12 billion valuation, led by Andreessen Horowitz, with Nvidia, Accel, Cisco, and AMD among the participants. The company confirmed its founding team to WIRED: CEO Mira Murati and cofounders John Schulman, Barrett Zoph, Lilian Weng, Andrew Tulloch, and Luke Metz. Murati had left her role as OpenAI’s chief technology officer the preceding September; the other cofounders brought experience spanning ChatGPT, research leadership, safety, robotics, pretraining, reasoning, and post-training.[11]

In September, Horace He, working with others at Thinking Machines, published Defeating Nondeterminism in LLM Inference. The work distinguished repeatable kernel execution from repeatable results for an individual user: changing server load changes batch sizes, which can change floating-point reduction order and ultimately generated tokens. In the team’s experiment, 1,000 temperature-zero, 1,000-token completions from Qwen3-235B-A22B for the same Richard Feynman prompt produced 80 distinct outputs. With batch-invariant kernels enabled, all 1,000 were identical.[10]

John Schulman, in collaboration with others at Thinking Machines, followed with LoRA Without Regret on September 29. LoRA preserves the original model weights while learning updates through two smaller matrices, reducing the memory needed for training and allowing multiple adapters to share a base model. The team reported that LoRA matched full fine-tuning on the small-to-medium instruction and reasoning datasets it tested when adapter capacity and layer coverage were sufficient. Attention-only adaptation underperformed; datasets exceeding adapter capacity and some large-batch settings produced gaps. Its mathematical-reasoning RL experiments also found matching performance at very low ranks.[9]

Thinking Machines launched Tinker in private beta on October 1, 2025. Researchers controlled their algorithms and data while the company handled distributed training on its internal clusters, including scheduling, resource allocation, and failure recovery. Tinker used LoRA to share compute across training runs. Low-level primitives such as forward_backward and sample let users express post-training methods; the open-source Tinker Cookbook supplied implementations on top of the API.[3]

The company’s launch announcement credited Princeton’s Goedel Team with training mathematical theorem provers and Stanford’s Rotskoff Chemistry group with fine-tuning for chemistry reasoning. Berkeley’s SkyRL group experimented with asynchronous off-policy RL involving multiple agents and multi-turn tool use. Redwood Research used Tinker to reinforcement-learn Qwen3-32B on difficult AI control tasks.[3]

On December 12, 2025, Tinker removed its waitlist and added Kimi K2 Thinking, OpenAI API-compatible sampling, and two Qwen3-VL vision models. The sampling interface could address checkpoints while training continued. Interleaved image and text inputs extended supervised fine-tuning and RL to pictures, screenshots, and diagrams.[4]

On March 10, 2026, Thinking Machines and NVIDIA announced a multi-year partnership to deploy at least one gigawatt of Vera Rubin systems, targeting early 2027. The agreement also covered work on training and serving systems for NVIDIA architectures. NVIDIA had made an investment in Thinking Machines, whose amount was not disclosed. The gigawatt figure described planned deployment, rather than installed capacity.[8]

Thinking Machines announced its interaction-model research preview on May 11. The system separated a continuously present interaction model from an asynchronous background model for sustained reasoning and tool use. Both shared conversation context, and background results streamed back into the ongoing exchange. TML-Interaction-Small was a 276-billion-parameter mixture-of-experts model with 12 billion active parameters.[5]

The interaction model interleaved 200 milliseconds of input with 200 milliseconds of output, preserving silence, overlap, and interruptions in its context. This supported simultaneous speech, timed interjections, and responses to visual cues without a separate dialogue-management component. The chunk duration was distinct from response latency: Thinking Machines reported 0.40-second turn-taking latency on FD-bench V1.[5]

Its encoder-free design used lightweight audio embeddings, 40×40-pixel image patches encoded by an hMLP, and a flow head for audio output, all co-trained with the transformer. Persistent streaming sessions appended incoming chunks to a sequence in GPU memory, reducing repeated allocation and metadata work; the company upstreamed a version to SGLang. Batch-invariant attention work involved Colfax. At the preview, long-session context management and reliable connectivity remained limitations, and larger pretrained models were too slow for this serving setting.[5]

On July 15, 2026, Thinking Machines released Inkling’s full weights. The mixture-of-experts transformer had 975 billion total parameters and 41 billion active, with pretraining on 45 trillion tokens of text, images, audio, and video. Its context window supported up to one million tokens; the launch offering on Tinker separately provided 64K and 256K options. Inkling reasoned over text, images, and audio with controllable thinking effort. The company positioned it as a broad base for customization and explicitly said it was not the strongest overall available model. One design goal was to serve as the background reasoning model in the interaction system.[6]

Inkling’s architecture largely followed DeepSeek-V3’s MoE design, while combining sliding-window and global attention at a 5:1 ratio, relative positional embeddings, and short convolutions. Training on NVIDIA GB300 NVL72 systems used Muon for large matrix weights and Adam for other parameters. Initial supervised post-training used synthetic data from open-weight models including Kimi K2.5; most post-training compute went to large-scale RL. Thinking Machines reported more than 30 million RL rollouts across two long continuous runs.[6]

A company demonstration made customization concrete: running inside OpenCode, Inkling generated training data and an evaluation, used Tinker to fine-tune toward responses containing no letter “e,” then loaded the resulting weights. In another demonstration, Inkling refined a multiplayer snake game through 40 iterations of feedback from GPT Codex as reviewer.[6]

Inkling-Small’s full-weight release followed on July 30, after its July 15 preview. It had 276 billion total parameters and 12 billion active, native audio and image reasoning, variable thinking effort, and up-to-one-million-token context. Beginning training later allowed revisions to its pretraining data and recipe. Its preview checkpoint used some on-policy distillation with Inkling as teacher, followed by two further weeks of agentic coding RL. Thinking Machines reported improvements over Inkling on reasoning and agentic coding benchmarks, while the larger model retained advantages in knowledge coverage and factuality. Both were available for fine-tuning on Tinker and chat through Tinker Playground.[6][7]