Thinking Machines Lab Ships Inkling 2 Small, a 594B Model Built for Long Jobs

Inkling 2 Small Is Here, and It Wants the Long Jobs

Thinking Machines Lab launched Inkling 2 Small on November 5, and the pitch is refreshingly specific. This is not another chatbot model trying to be good at everything. It is a 594-billion-parameter system built for long-horizon autonomous work, the kind of multi-hour coding, research, and agentic tasks where most models quietly lose the plot halfway through.

The model is a sparse mixture-of-experts design: 594 billion total parameters, but only 17 billion active for any given token. A 48-layer decoder-only transformer routes each token to eight of 384 experts, mixing local and global attention. That is the architecture of a lab that cares more about sustained reasoning than about winning a benchmark by half a point.

What You Can Actually Use Today

At launch, Inkling 2 Small is available through Tinker, the lab's platform, with inference-only access in the Tinker Playground and fine-tuning plus sampling under Tinker's terms. There is no download yet, and that is deliberate.

Thinking Machines Lab says it intends to publish the full weights under Apache 2.0 in the week following the announcement, but only after external safety testing and red-teaming are complete. Fine-tuned checkpoints would ship with those weights, and more platforms are planned after the safety work finishes. The company is explicit that publishing weights is irreversible, so the gate is real, not theater.

This follows the lab's staged-release framework from July 31, 2026, which widens access step by step: inference first, then monitored hosted fine-tuning, then open weights. Progression depends on evidence, not on a calendar. If the external testing finds problems, the weights wait.

Why Long-Horizon Work Is the Hard Part

Most model launches this year have been about speed, price, or a bigger context window. Inkling 2 Small is aimed at a different bottleneck. Agents that work for hours need to keep track of what they already tried, recover from their own mistakes, and stay on task when the instructions get messy. That is still the weakest part of the current generation, and the lab knows it. The AI agent wars are increasingly fought on exactly this ground: who can keep an agent useful for a whole afternoon, not just a clever first reply.

The timing matters too. The open-weight field has been moving fast this week, with Tencent shipping Hy4 and Moonshot putting Kimi K3.1 Pro weights on Hugging Face. Inkling 2 Small arrives into a crowded race, but with a clearer identity than most: a big model, tuned for autonomy, from a lab that has been careful about how it opens things up.

The Open-Weights Promise, With a Catch

Open weights under Apache 2.0 would make Inkling 2 Small one of the most capable downloadable models around, and developers should pay attention. A 594B model with 17B active parameters is the kind of system that can run on serious but not absurd hardware, which is exactly the sweet spot for teams that want to fine-tune without renting a hyperscaler's entire fleet.

But the conditional part matters. The weights arrive only if the safety testing passes, and the lab has not published a date. Treat the promise as a plan, not a guarantee. If you are building on Inkling 2 Small today, build on the hosted version and treat the open release as a bonus.

What to Watch Next

The real test is not the launch post. It is the external evaluation results, the weight release, and whether developers actually pick this model up for the long jobs it claims to handle. Meta's Muse Spark showed this year that a huge open-weight model can make a splash, and Gemini 4 Argon showed that hosted frontier models keep raising the bar. Inkling 2 Small sits between those two stories: hosted now, open soon, and judged on whether it can stay coherent when the task takes hours instead of seconds.

If the weights land next week as planned, this becomes one of the most interesting open releases of the year. Until then, it is a promising hosted model with a homework assignment still to finish.