Thinking Machines’ Inkling Takes U.S. Open-Model Lead With 41 Score

Artificial Analysis independently scored Inkling at 41 on its Intelligence Index, putting Thinking Machines Lab’s first production model ahead of every other U.S. open-weight release the firm tracks. Thinking Machines released Inkling on Wednesday, July 15, with downloadable weights that developers can run and adapt. Chinese models remain ahead on several coding and reasoning tests in the lab’s own table, which confines the 41 result to the U.S. portion of the index. The release exposes model checkpoints, while the source code and exact training corpus remain undisclosed.

Inkling’s 41-Point Intelligence Index Result

The benchmark firm put Inkling three points ahead of Nemotron 3 Ultra, the previous U.S. open-weight leader. Its v4.1 index combines nine evaluations, including agentic work, coding, scientific reasoning, knowledge reliability and long-context performance. The published methodology describes those evaluations as independently measured. Several of the same benchmarks also appear in Thinking Machines’ launch table, which mixes outside scores with results from the lab’s own harnesses.

Inkling reached an Elo of 1,238 on GDPval-AA v2, ahead of Kimi K2.6 at 1,190 and DeepSeek V4 Flash max at 1,189. It scored 24% on Tau 3 Banking, compared with 21% for Kimi and 23% for DeepSeek.

The firm’s model page recorded 72.7 output tokens per second through Thinking Machines’ API and 1.75 seconds to the first token. At the 64K context tier, the service listed prices of $1.87 per million input tokens and $4.68 per million output tokens. Those prices compare with class averages of $0.43 and $1.25, respectively. At the 256K tier, the launch analysis listed $3.74 per million input tokens and $9.36 per million output tokens. Provider prices can differ.

On AA-Omniscience, Inkling registered 40% accuracy and a 63% hallucination rate. That test rewards correct answers, penalizes unsupported answers and leaves refusals unpenalized, so the rate applies to one knowledge benchmark rather than every Inkling workload. Thinking Machines’ model card advises developers to verify outputs used in high-stakes settings.

Inkling averaged 25,000 output tokens per Intelligence Index task, compared with GLM-5.2 max at 43,000, Kimi K2.6 at 38,000 and DeepSeek V4 Pro max at 37,000. Across the broader comparison class, however, Inkling generated 130 million tokens during the index run, above the 92 million average. The benchmark firm consequently described it as somewhat verbose against that wider group.

Inkling’s Scores Against GLM-5.2 and Gemini 3.1 Pro

Thinking Machines opens its launch report with an explicit limit, writing that Inkling is “not the strongest overall model available today, open or closed.” The company positions it as a model that organizations can fine-tune, with text, image and audio inputs plus a control for the amount of reasoning used on each task.

The company’s table combines outside benchmark scores with results from its own harnesses. Thinking Machines used scores reported by Artificial Analysis for Humanity’s Last Exam, GPQA Diamond, GDPVal, Tau 3 Banking and AA-Omniscience, among other rows. The lab reported 77.6% on SWE-bench Verified, compared with 70.7% for Nemotron 3 Ultra. Inkling’s 63.8% on Terminal Bench 2.1 remained below GLM-5.2 at 82.7%, while its 29.7% on the text-only Humanity’s Last Exam trailed GLM-5.2 at 40.1%. On audio, the company put Inkling at 91.4% on VoiceBench and Gemini 3.1 Pro at 94.3%.

The table below separates the two sources of those numbers.

Thinking Machines said all coding evaluations used a 256K maximum-token trajectory limit. The lab used a bash-only harness for its SWE-bench result and an internal coding harness for Terminal Bench, assigning zero when a solution showed contamination from web search. It used self-reported results for outside models where external scores were unavailable. The model card repeats the contamination rule and lists the benchmark table.

Inkling’s effort control is a system setting available to developers. Thinking Machines swept that setting from 0.2 to 0.99 and said Inkling matched Nemotron 3 Ultra on Terminal Bench with roughly one-third as many generated tokens. That comparison came from the lab itself. The benchmark firm’s 25,000-token test offers an independent measure of Inkling’s output consumption.

The launch materials also describe the model’s post-training process. Inkling’s initial supervised fine-tuning relied on synthetic data generated by open-weight models, including Kimi K2.5. Thinking Machines says this phase accounted for only a small share of the total training compute. It was followed by large-scale reinforcement learning in synthetic and human-created environments. The company does not disclose how much of the fine-tuning data came from Kimi K2.5 or how that data affected the released checkpoint.

The model card also summarizes the company’s safety findings. In internal and external evaluations, Inkling scored below publicly available frontier models on strategic deception, sabotage and other measures associated with loss of control. However, the lab found that the model sometimes complied with harmful requests when they were presented as role-play or phrased indirectly. For consumer-facing or high-traffic applications, Thinking Machines recommends adding input and output classifiers such as Llama Guard.

Inkling’s Two-Terabyte Hardware Floor

Running Inkling’s BF16 checkpoint requires at least two terabytes of combined GPU memory, according to the model card. Thinking Machines lists eight Nvidia B300 GPUs or 16 H200 GPUs as suitable configurations. The NVFP4 checkpoint reduces the minimum requirement to 600 gigabytes. Supported configurations include four B300 GPUs in W4A4 mode or eight H200 GPUs in W4A16 mode.

Behind those requirements is a 66-layer mixture-of-experts transformer with 975 billion parameters in total, of which 41 billion are active for each token. Thinking Machines says the model was pretrained on 45 trillion tokens spanning text, images, audio and video. The released checkpoint accepts text, image and audio inputs, generates text, supports a context window of up to one million tokens and is available under the Apache 2.0 license.

The model’s sparse architecture sends each token to six of its 256 experts while keeping two shared experts active. This design explains the large difference between the model’s total and active parameter counts. Thinking Machines uses five local-attention layers for every global-attention layer. Images are processed through a hierarchical patch encoder, while audio is converted into discrete dMel representations. The decoder then handles all input types within a shared hidden space.

What that architecture costs to run, and to rent, is set out below.

According to the model card, the training data came from public sources, licensed or otherwise acquired third-party material, and synthetic or augmented data. It describes cleaning, deduplication and filtering without naming the individual datasets used for the 45-trillion-token run.

Hugging Face’s launch review reported release-day support in Transformers, SGLang, vLLM and llama.cpp. Its reviewers labeled their small hands-on image and audio checks “vibe evaluations,” rather than a general benchmark. Inkling answered all five selected image questions at high effort and missed one at medium effort; it passed the chosen audio questions at medium effort and missed one formal-fallacy example at the lowest setting.

The lab had already signed a multi-billion-dollar Google Cloud agreement for Nvidia-based AI infrastructure. The model card now gives outside operators the cluster sizes needed to deploy the same open-weight checkpoint through their own systems.

Tinker’s Paid Fine-Tuning Model

Thinking Machines made Inkling’s weights free to download and sells Tinker as a paid fine-tuning service. TechCrunch reported that training, fine-tuning and participation in the hosting market form the company’s revenue path around Inkling. Organizations that download the weights do not owe Thinking Machines a metered model-access fee, although running the smallest model-card configuration still requires four B300 GPUs.

Futurum Group analyst Mitch Ashley told The Wall Street Journal that “Engineering teams should treat base-model selection as an architecture decision.” Ashley linked each downstream customization to the work required if an organization later switches its base model.

Constellation Research analyst Holger Mueller told SiliconANGLE that Tinker may be the company’s larger business-model innovation because Thinking Machines charges customers to customize the model instead of charging for access to its weights.

The model card leaves deployment safeguards with the operator. It asks downstream developers and deployers to test performance, safety and fairness for their use case, add content filtering, rate limits and monitoring, and keep human review for medical, legal or safety-critical work. Thinking Machines also says Inkling can falter during long multi-turn conversations and perform unevenly across languages or subject areas that received less training data. Those responsibilities apply before and during deployment.

Fine-tuning requires specialized machine-learning staff, TechCrunch noted. The outlet cited a Bridgewater Associates project as evidence for Tinker’s value proposition. Thinking Machines and Bridgewater said a customized open model achieved 84.7 percent on financial reasoning tests while costing about one-fourteenth as much to run as leading proprietary systems.

The project, however, used a different base model. The result also came from a joint evaluation by the two companies, not from an independent test of Inkling.

Inkling is available through Tinker, as downloadable checkpoints, and through APIs offered by Together AI, Fireworks, Modal, Databricks and Baseten. Tinker provides context-window options of 64,000 and 256,000 tokens, while the open-weight checkpoints support up to one million tokens.

Thinking Machines also lists SGLang, vLLM, TokenSpeed, llama.cpp and Transformers as supported deployment tools.

The company has separately released a preview of Inkling-Small, which has 276 billion total parameters and activates 12 billion for each token. Thinking Machines says it is completing its tests and will publish the smaller model’s full weights once that work is finished.

GLM-5.2 Edges Kimi K2.7 Code in Early Coding Tests

The two strongest Chinese open models to launch this month, Z.ai's GLM-5.2 and Moonshot's Kimi K2.7 Code, have now been run side by side by five independent reviewers, and the early verdict gives GLM-

The Implicator

MiniMax promises M3 weights after 1M-context model launch

MiniMax released M3 on Monday and said model weights and a technical report will follow within 10 days, leaving developers with API access before local inspection. The Shanghai company is offering M3

The Implicator

Nvidia's Nemotron 3 Ultra Leads US Open Models but Trails China's Kimi K2.6

Nvidia unveiled Nemotron 3 Ultra at its GTC Taipei keynote on Monday and released a benchmark, run with the evaluation firm Artificial Analysis, that places the model first among open-weight systems b

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai