Inkling 2 Small: Murati's lab launches a 594B model, but open weights will have to wait

What's going on

Thinking Machines Lab unveiled Inkling 2 Small on November 5, 2026. The San Francisco startup was founded by Mira Murati, OpenAI's former CTO. Inkling 2 Small is the lab's second model generation, arriving less than four months after the first Inkling family in July. It's designed for long tasks that it completes on its own (in industry jargon, it's agentic: the model plans and carries out multiple steps by itself instead of just answering a question). It's also meant to help with coding, vision tasks and knowledge work.

If you want to run the model yourself, there's a catch: for now, Inkling 2 Small is closed-weights only and available through Tinker, the lab's platform. In the Tinker Playground you can only run the model (inference). Through Tinker itself, you can also fine-tune it and generate outputs. Thinking Machines plans to release the full weights under the permissive Apache 2.0 license. According to Unite.AI, that's planned for the week after the announcement, but only once outside safety testing and red-teaming are complete. Red-teaming means experts deliberately try to get the model to behave harmfully. Crypto Briefing, on the other hand, stresses that there's no firm date. Mira Murati only wrote on X that the weights are coming "soon." If you build your own fine-tune on Tinker, you should be able to download it along with the weights. Other platforms are supposed to follow after the safety testing.

Architecture: lots of knowledge, little compute

Inkling 2 Small is a Mixture-of-Experts model. Think of it as a big team of specialists where only a few get called in for each subtask. The model has 594 billion parameters in total, but only 17 billion are active per token. According to the Model Card, it's a decoder-only transformer with 48 layers. Each token is routed to 8 of 384 experts. It combines local and global attention, so the model looks partly at the nearby text and partly at the entire context.

The model takes text and images as input, processes both in a shared internal space and responds in text. Unlike the large original Inkling, it doesn't support audio. The context window is 1 million tokens, meaning the amount of text the model can keep in view at once. It supports the MXFP8 and NVFP4 number formats, which store numbers in 8 and 4 bits respectively and save memory and compute. You can also adjust how hard the model thinks. More effort costs more time and money but gets you more thorough answers.

In terms of size, the new model sits between its predecessors:

So "Small" is relative: Inkling 2 Small is more than twice the size of Inkling-Small.

How it was trained

Thinking Machines is unusually open about its training details. Pretraining covered 8.4 trillion tokens of language, code and images. It took 68 hours on Nvidia's GB300 NVL72 racks and ran natively in the 4-bit NVFP4 format. Further phases followed:

• 3.8 trillion tokens of newer, higher-quality data at 512,000 tokens of context,
• around 320 billion tokens of reasoning and agent trajectories at 1 million tokens of context,
• about 34 billion tokens of curated agent trajectories.

After that came one large joint reinforcement learning run (the model tries out solutions and gets rewarded for good results). According to the company, it covered 1,075 steps and more than 17 million rollouts. It took nine days on 2,048 GPUs and used over 3 trillion tokens. Coding made up 38.5 percent of rollouts and 65 percent of tokens. Other areas included knowledge work, math and science, web and app building, image understanding, tool use, search, chat and safety. The context length was 256,000 tokens, supplemented by compaction: older parts of a trajectory get summarized to make room for new ones. The average rollout took eight minutes, and the longest (99th percentile) took four hours. Randomly assigned effort levels and a penalty for overly long answers were meant to teach the model to balance speed against thoroughness. According to Crypto Briefing, the model also uses visual feedback to correct its work when building websites and games.

The company showed off how independently the model can work with a demo. Inkling 2 Small was tasked with training the smaller Qwen3.5-4B model on time-series forecasting via Tinker. To do that, it built its own evaluation routine and created 12,500 training tasks. It tried supervised fine-tuning and reinforcement learning and debugged on its own. Along the way, it found a bug in how the evaluation's random seed was initialized. When reinforcement learning made the forecasts worse, it abandoned that approach. After 4 hours and 17 minutes, it submitted its best checkpoint, having spent $426 of a $500 budget. According to the company, the result made about 15 percent fewer errors on unseen data than the base model.

Benchmarks and pricing

Inkling 2 Small scores 31 points on the Artificial Analysis Intelligence Index v5.0. Thinking Machines calls that a "big leap" over its predecessor. On DesignArena, a leaderboard for design and frontend tasks, it ranks 4th overall and 1st among open-weight models, according to Crypto Briefing. That label is a bit odd as long as the weights aren't actually available yet. On Tinker, the model costs $0.30 per million input tokens and $1.00 per million output tokens.

The training incident

The touchiest part of the announcement is about safety. Thinking Machines admits that early checkpoints sometimes made up results, ignored instructions and covered up what they were doing once tasks got hard. Midway through the reinforcement learning run, a monitoring system raised an alarm: the model had failed at a task and, in the process, got around a network block through a third-party service it could reach with its search tools.

The company paused the run and locked down its infrastructure. It removed environments with internet access, and every tool call is now checked before it runs. Training then continued with additional alignment phases, meaning training aimed at the desired behavior. Other measures:

• Training data from the intermediate phase was filtered.
• A general reward signal for good behavior was added.
• Dedicated training environments were built for specific failure patterns.
• The model also learned from a teacher model guided by a behavioral spec (self-distillation).

Monitoring now checks tool calls against task-specific policies and fixed rules. When in doubt, another language model acts as the judge. A second reviewer checks completed rollouts after the fact.

The safety testing covered everyday use, plus chemical, biological, radiological and nuclear weapons risks, cyberattacks, variants with refusals switched off, deliberate deception and sabotage. According to the company, the model is well below frontier model capabilities when it comes to loss-of-control risk. It's said to be better aligned than earlier Inkling models, but worse than GPT-6.1 Sol in the company's own tests.

The release follows a staged plan the lab presented on July 31, 2026. First comes inference only, then supervised fine-tuning on the company's own platform, and possibly open weights after that. In the fine-tuning stage, users can bring their own data and training methods, but the provider monitors usage and can cut off access. That step depends on evidence and doesn't happen automatically. The company stresses that releasing weights can't be undone.

Known weaknesses

The Model Card is upfront about a range of limitations:

• no audio support,
• a tendency toward rambling answers,
• unreliable tool selection in codebases with many skills,
• weaker structure and formatting in chat,
• overreactions to some questions about emotional support and safety,
• hallucinations, meaning made-up facts,
• imprecise instruction following,
• declining quality in long conversations,
• uneven performance in less common languages and specialist fields.

For high-stakes use, the company recommends that humans review the outputs.

Our take

Thinking Machines was founded in February 2025 with $2 billion in seed funding at a $12 billion valuation. The lab is betting on customizable models rather than chasing the title of strongest chatbot. It explicitly marketed the first Inkling as "not the strongest model," but as a good open base for fine-tuning. Inkling 2 Small continues that approach, with low prices and strengths in coding and design.

Crypto Briefing points out an interesting contradiction: the lab presents itself as a champion of open weights, yet its new model ships closed at first. The training incident does make that caution understandable, though. The very autonomy the model is marketed on is also what led it to get around a block. With open weights, the model would leave the environment where Thinking Machines can monitor usage and cut off access. It's still unclear when, or whether, the release will actually happen. The benchmark scores and safety results so far also come mostly from the company itself or from leaderboards. Independent testing is still to come. And even once it's released, a 594-billion-parameter model is nothing you'll run on a regular PC. Even heavily compressed (4-bit), it would need roughly 300 GB of memory or more.