Key Notes
- Thinking Machines has released Inkling 2 Small on Tinker for visual coding, tool use and extended autonomous work, with inference and fine-tuning access available now.
- The lab reports that an agent trained a separate forecasting model in a few hours and more than doubled its internal benchmark score, while its website leaderboard position remains close to competing models.
- Full weights are planned under Apache’s permissive license after external safety testing, and training charges are separate from input and output token prices.
Thinking Machines Lab has released Inkling 2 Small, a model designed to write code, inspect visual results and keep working through extended tasks. It is available on the company’s Tinker platform for inference and fine-tuning, with downloadable weights planned for the following week after external safety testing.
The November 5 announcement puts visual coding and autonomous research at the center of the release. Rather than stopping after producing a program, the model can use images of its output to check and revise websites, games and interactive 3D scenes.
Visual Feedback Becomes Part of the Coding Loop
Inkling 2 Small accepts text and images and produces text, according to its model card. Its visual creations come through tools and code: the launch demonstrations include a procedurally generated city simulator built with Three.js and a Blender scene featuring a paper landscape and origami cranes.
The architecture uses a mixture of experts, with 594 billion parameters in total and 17 billion active for each token. For each token, the network selects a subset of its experts while retaining the much larger total parameter pool. The model also supports adjustable reasoning effort, allowing developers to trade a longer run against the depth of work required.
For a coding agent, the practical benefit of visual input is the chance to catch problems that a successful code execution might miss. A site can render with a poor layout, and an interactive scene can look wrong even when the program runs. Giving the agent a view of the result lets it include those problems in its next round of changes.
The Website Benchmark Lead Has Already Moved
Thinking Machines’ November 5 DesignArena Website snapshot put Inkling 2 Small at 1340, above GPT-6 Astra at 1333 and Claude Opus 5.5 at 1321. Those are models from OpenAI and Anthropic, respectively. The comparison concerns website output on a particular leaderboard, rather than an overall ranking of coding or reasoning ability.
By November 6, the live leaderboard showed Inkling at 1334, just behind Astra at 1335 and still above Opus at 1326. The narrow margins and changing positions are useful context for the launch claim. Separately, Artificial Analysis lists an Intelligence Index score of 31 for Inkling 2 Small.
An Agent Training Another Model
The release’s research demonstration asked Inkling to adapt Qwen3.5-4B for time-series forecasting, including weather and Wikipedia page views. The agent analyzed the task, wrote an evaluator, selected data, experimented with supervised and reinforcement learning, and debugged the training process. Thinking Machines reports that it worked for 4.5 hours and incurred $426 in Tinker charges.
The company says the resulting forecasting score more than doubled. Its replay shows a held-out comparison improving from 10% less error than a simple copying baseline to 23% less error. That measures improvement against a baseline in this internal experiment; it is not a claim that forecasting accuracy doubled across arbitrary real-world problems.
The work illustrates a different role for a coding model: coordinating a sequence of experiments, inspecting the results and deciding what to try next. Its usefulness depends on the evaluator and data as well as the agent’s ability to keep the workflow running.
Tinker Pricing and the Planned Weight Release
The current Tinker pricing lists $0.30 per million input tokens and $1 per million output tokens, with cached input at $0.03. Training is charged separately: the listed rate is $1.50 per million training tokens for the 64K configuration and $3 for the 256K version. Serverless inference remains in beta, with production access handled through a waitlist.
Users can try the model in Tinker Playground now. Thinking Machines intends to release the full weights under Apache 2.0 after external testing and red-teaming, so the open-weight release is still a plan at publication. The company also lists limitations, including verbose communication, imperfect tool choices and no audio support.
We previously examined the lab’s interaction models, which focused on continuous collaboration across audio, video and text. Inkling 2 Small extends the company’s product range toward visual coding and longer independent workflows, with fine-tuning available to developers who want to adapt it to their own tasks.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
