The most interesting part of this release is not that Thinking Machines has produced a new frontier-model contender, but that the video argues it was never really trying to. After establishing the extraordinary expectations surrounding Mira Murati’s departure from OpenAI, the company’s reported $12 billion valuation, and the talent assembled around her, the presentation pivots toward a less glamorous proposition: Incline can be comparatively ordinary as a general-purpose model because its value is supposed to emerge through efficiency, openness, and customization. That tension between enormous expectations and a deliberately non-dominant model gives the episode a stronger angle than another benchmark roundup.
The technical overview packs a surprising amount into a short runtime. Incline is described as a 970-billion-parameter mixture-of-experts model activating 41 billion parameters per token, pretrained on 45 trillion tokens spanning text, images, and audio, with a one-million-token context window and openly available weights. The explanation of mixture-of-experts routing is particularly effective because it communicates why total parameter count and active computation are different without getting buried in architecture details. The video also emphasizes direct processing of pixels and audio rather than separate encoders, though the rapid delivery leaves little room to examine what practical advantages that design produces.
Crucially, the video does not pretend the model dominates conventional benchmarks. Incline is characterized as landing well behind leading proprietary systems and around the middle of competing open models, with the timing made more awkward by Moonshot's much larger Kimi release immediately afterward. Those comparisons are presented with intentionally unserious labels such as “Trust Me Bro Benchmarks,” which fits The Code Report's comedic style but also signals an important limitation: viewers are given summarized performance claims rather than enough methodology, benchmark context, or underlying results to independently judge how meaningful the gaps are.
The stronger case for Incline comes from features intended to matter beyond leaderboard position. “Thinking Effort” is presented as a way to trade reasoning expenditure for speed and cost, with the claim that higher settings can match Nemotron 3 Ultra on Terminal Bench while consuming roughly one-third as many tokens. That could be significant for agents executed at enormous scale, where inference efficiency compounds quickly. The video also highlights training intended to reward uncertainty awareness, claiming this makes Incline particularly strong at forecasting future events and capable of outperforming GPT-5.5 and Opus 4.8 in that area. These are interesting results as presented, but they remain claims within the video rather than established conclusions because the supporting evaluations are not explored in enough detail to validate them.
The self-modification demonstration is the episode's most memorable example. Connecting Incline to Tinker allows the model to write a training script, create data, run fine-tuning, and load modified weights that remove its ability to use the letter E. It is deliberately absurd, but it illustrates the intended relationship between the open model and Thinking Machines' fine-tuning infrastructure more clearly than abstract benchmark numbers could. The discussion of reinforcement learning causing the model's internal reasoning to become increasingly compressed is similarly intriguing, although the caveman-English joke gets more attention than the implications of that behavior.
That leads to the video's most convincing interpretation of Thinking Machines' business strategy: Incline is the accessible foundation, while Tinker is where customers can turn that foundation into specialized models and where the company potentially captures value. Framing a supposedly “mid” general model as raw material for specialists makes the release much easier to understand in the context of the company's investment and research talent. It is a persuasive strategic reading, but still a reading; the video moves quickly from technical characteristics to a confident explanation of the company's commercial play without providing much evidence about adoption, pricing, customer demand, or whether specialized fine-tuning will actually deliver the suggested advantage.
Presentation is unmistakably The Code Report: extremely fast, dense with technical terminology, and constantly punctuated by jokes about Silicon Valley money, benchmark competition, pop culture, and model behavior. The humor keeps a potentially dry model release entertaining, and several jokes double as effective explanations, but the pace sometimes works against the subject. Claims involving forecasting performance, token efficiency, multimodal architecture, and training behavior deserve more scrutiny than they receive. The Clerk sponsorship also occupies a substantial portion of the ending and arrives immediately after the video's strategic conclusion, making the substantive analysis feel as though it stops just as the most important business questions become interesting.
Pros
- Reframes Incline around efficiency, openness, and specialization rather than treating benchmark leadership as the only meaningful measure of a model.
- Explains the 970-billion-parameter mixture-of-experts architecture and 41-billion active-parameter figure in an accessible way.
- The self-modification experiment with Tinker provides a memorable demonstration of the model's intended fine-tuning workflow.
- Acknowledges weak comparative benchmark positioning instead of presenting the release as an unquestioned frontier breakthrough.
- Connects Thinking Effort and reduced token usage to a practical concern for large-scale agent workloads.
Cons
- Major performance claims, including forecasting results and Terminal Bench efficiency, receive too little methodological context to judge their significance.
- The rapid pace compresses several technically important subjects into brief explanations before moving to the next joke or feature.
- The interpretation of Tinker-centered specialization as Thinking Machines' real commercial strategy is plausible but presented more confidently than the evidence shown supports.
- Little attention is given to adoption, pricing, deployment requirements, or other practical factors that would help determine whether the model-and-fine-tuning strategy is commercially compelling.
- The extended Clerk sponsorship makes the ending feel abrupt after the core analysis establishes questions worth exploring further.
This is an entertaining and unusually clear explanation of why a heavily funded AI company might deliberately release an open model that does not dominate the leaderboard, with Incline's adjustable reasoning, unusual multimodal approach, and integration with Tinker providing a coherent alternative to pure benchmark competition. The strongest demonstrations make specialization feel like a credible direction, but the video moves too quickly through performance evidence and business assumptions to establish that Thinking Machines has already validated the strategy it describes.
