TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Thinking Machines says Inkling 2 Small has made large benchmark gains over its predecessor, particularly on agentic coding and visual tasks. The model is available through the company’s Tinker platform; its weights are planned for release under Apache 2.0 after external safety testing.
Thinking Machines, the AI startup founded by former OpenAI chief technology officer Mira Murati, has introduced Inkling 2 Small, a multimodal model it says ranks among the strongest Western open-weight models on selected benchmarks. The model is currently offered through the company’s Tinker platform, while its weights are slated for release under an Apache 2.0 license after external safety testing.
The model uses a mixture-of-experts architecture with 594 billion parameters in total and 17 billion active for a given computation, according to the report. It takes text and images as input and includes adjustable thinking depth, allowing users to trade faster responses for more extensive reasoning. Thinking Machines describes Inkling models as base models intended for fine-tuning, rather than systems optimized primarily for general chat.
On the Artificial Analysis Intelligence Index, Inkling 2 Small scores 31 points, compared with 13 for its predecessor, Inkling Small. The report says that result is roughly on par with Google’s Gemini 3.8 Flash. Gains are particularly pronounced on agentic benchmarks: the model’s AutomationBench-AA score rises from 5 to 46 percent, while its FrontierCode 2.0 score rises from 5 to 25 percent. On AA-Briefcase, a knowledge-work benchmark, it records 1,527 Elo, 615 points above the prior model.
Those results come with trade-offs. Artificial Analysis reports an average of about 150,000 output tokens per task, roughly four times the predecessor’s usage, and a cost of $1.18 per task, nearly six times higher than before. The reported cost remains below Gemini 3.8 Flash at $7.77 and Claude Sonnet 5.5 at $10.02. The model’s hallucination rate falls from 61 to 27 percent, but the report says it also answers fewer questions: its response rate declines from 77 to 53 percent.
Coding Gains Come With Higher Usage
Inkling 2 Small’s results matter because they suggest a new contender for developers who want a model they can adapt to specialized tasks, particularly coding and visual analysis. Its reported gains on multi-step agentic work are substantial compared with the predecessor, and the company says it achieved them using a training run of about 12 trillion tokens across all phases.
However, benchmark leadership is not the same as proven superiority across everyday applications. The model’s higher token use and cost per task may affect deployment economics, while its lower response rate complicates the improved hallucination figure. And until the weights are released, developers cannot independently run or inspect the model in the way an open-weight release would allow.
For coding comparisons, the report also identifies strong competition. Inkling 2 Small scores 25.4 percent on FrontierCode 2.0 and 72.1 percent on DeepSWE, while DeepSeek V4.1 Flash leads those particular results at 28.5 percent and 77 percent, respectively. On SWEBench Pro v2 (Hard), Kimi K3 is reported to lead narrowly. These results make the model’s claimed top spot dependent on which benchmarks and competitors are included.
Products Worth Considering
Bose QuietComfort Ultra Earbuds (2nd Gen) – Wireless Noise‑Cancelling Earbuds with Mic, Immersive Audio, USB‑C Charging, Up to 6 Hours Batte
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, 12MP Center Stage Camera, Touch ID, Wi-Fi 7, Midnight
Kindle Paperwhite Signature Edition 32GB (2024) – Auto-adjusting front light, wireless charging, weeks of battery life – Metallic Black
As an affiliate, we earn on qualifying purchases.
A Base Model Before Weight Release
Thinking Machines presents the Inkling family as a set of base models for fine-tuning, which can be useful for organizations building systems for narrower tasks. That positioning differs from models chiefly tuned for conversational use. The company says the new model was trained with reinforcement learning on hundreds of thousands of long-horizon coding tasks, using an automated evaluation system to check whether generated code worked and how the model approached a task.
The report gives additional results across coding and visual tests: Inkling 2 Small reaches 86.3 percent on SWEBench Pro v2 (Hard), 25.2 percent on GDP.pdf, 81.6 percent on MMMU Pro and 64.7 percent on PerceptionBench. Thinking Machines also reports a 1,340 Elo rating on DesignArena Website, the highest among open-weight models in the comparison. These are benchmark-specific results, not a single overall measure of performance.
The model has limitations acknowledged in the report. It does not support audio, is less optimized for chat, may communicate inefficiently, and does not always select the best tools for large codebases. The company also says it can overreact to health-related topics and produce cluttered output. Its training process reportedly used 3 trillion tokens during the reinforcement-learning phase, run on 2,048 GPUs over nine days.
“The company describes Inkling 2 Small as its strongest Western open-weight model.”
— Thinking Machines
Rankings Await Public Access
The supplied report does not establish an independent, universal ranking of Western open-weight models. The “strongest” description is a company claim tied to selected benchmark results, and comparisons can shift with benchmark choice, model versions and evaluation methods. The report notes that newer Kimi K3.1 Pro is absent from the comparison.
It is also not yet clear when external safety testing will be completed or when the weights will be published. Until then, independent researchers and developers have limited ability to reproduce the reported results using the released weights. The benchmarks do not by themselves show how the model will perform in real deployments, or whether its higher token consumption will offset its lower stated task price.
Safety Review Precedes Public Weights
The next stated milestone is the completion of external safety testing, after which Thinking Machines plans to publish the weights under Apache 2.0. The model remains accessible through Tinker in the meantime. The company has also said that more sizes in the Inkling model family are planned, but has not provided release dates in the source material.
Readers and developers will be able to judge the benchmark claims more directly once the weights and further evaluation details are available. Until then, the performance figures, cost estimates and planned licensing should be treated as reported results and company plans, rather than independently verified outcomes.
Key Questions
What is Inkling 2 Small?
Inkling 2 Small is a text-and-image model from Thinking Machines, designed as a base model for fine-tuning. The company says it has 594 billion total parameters, with 17 billion active at a time.
Is Inkling 2 Small publicly available as open weights?
Not yet, according to the report. Thinking Machines currently offers it through Tinker and plans to release the weights under Apache 2.0 after external safety testing.
What benchmark results does the company highlight?
The reported results include a score of 31 on the Artificial Analysis Intelligence Index, 46 percent on AutomationBench-AA and 25 percent on FrontierCode 2.0. Scores vary by test, and competitors lead on some coding benchmarks.
What are the main limitations reported?
The model uses more output tokens per task than its predecessor and has a lower response rate, despite a lower reported hallucination rate. The report also says it lacks audio support, is less optimized for chat and may choose tools poorly for large codebases.
Source: the-decoder.com
