When Thinking Machines — the lab founded by former OpenAI CTO Mira Murati — released Inkling-Small on July 15, it did something no US-based lab had done since the open-weights wave began: ship a frontier-class model that competes directly with Chinese open-weights releases on both capability and price.
The model is a Mixture-of-Experts architecture with 276 billion total parameters but only 12 billion active at inference, a design that keeps compute costs low while preserving the reasoning depth of its larger sibling, Inkling (975B total, 41B active). On the benchmarks that matter for agentic automation — the kind of autonomous software workflows that are rapidly becoming the industry’s center of gravity — Inkling-Small performs at or near the level of models four times its size.
SWE-Bench Verified, the standard test for code-fixing agents, comes in at 80.2 percent. Terminal Bench 2.1, which measures autonomous command-line execution, scores 64.7 percent. AIME 2026, a math reasoning benchmark, reaches 95.1 percent at maximum effort settings. These are not abstract numbers. They represent the practical ceiling for what an open-weights model can do when deployed as a software engineering agent, a cybersecurity scanner, or a research assistant operating in long-horizon, multi-step workflows.
The pricing tells the strategic story. On Thinking Machines’ Tinker API, serverless inference runs $0.30 per million input tokens and $1.20 per million output tokens at 256K context — roughly half the cost of OpenAI’s Luna ($0.20/$1.20 for input/output, but behind a closed API) and well below Kimi K3’s $3.00/$15.00. DeepSeek’s V4-Flash-0731, the current pricing floor at $0.14/$0.28, remains cheaper, but Inkling-Small offers something DeepSeek does not: a 1-million-token context window and native multimodal support in a single open-weights package.
The two-model strategy mirrors what DeepSeek executed with Flash and Pro. Inkling (the larger model) handles the highest-complexity reasoning tasks; Inkling-Small takes the agentic and efficiency-sensitive workloads. This is not a stripped-down demo model. It is a deliberate product split designed to let developers deploy the smaller model for the majority of their agent pipelines while reserving the larger model for tasks that genuinely require deep reasoning.
The fine-tuning story matters as much as the inference pricing. Tinker, Thinking Machines’ training API, offers fine-tuning at $1.73 per million tokens for 64K context (currently at a 50 percent introductory discount). This is cheaper than self-hosting on most cloud GPU configurations and dramatically lower than the engineering cost of building a custom fine-tuning pipeline. The bet is that developers will not just use the model — they will specialize it, creating a moat of custom weights that makes switching costs real.
Adoption is still early. Hugging Face download data shows roughly 4,000 lifetime downloads as of July 31, a fraction of the volume Kimi K3 attracted in its first week. But the number is misleading as a measure of commercial traction. Thinking Machines is selling through Tinker, not primarily through Hugging Face downloads. The real adoption signal will come from enterprise deployment announcements and the emergence of specialized fine-tunes on downstream platforms.
The US-China dimension is unavoidable. Since DeepSeek and Moonshot began shipping capable open-weights models at aggressive price points, the narrative has been that US labs are losing the open-weights race. Inkling-Small does not end that narrative — DeepSeek’s pricing advantage is structural, driven by lower compute and labor costs — but it challenges the premise that US labs cannot or will not compete on openness.
Murati’s move from OpenAI, where she helped build the most commercially successful closed model in history, to a lab that ships open weights on Hugging Face is itself a signal about where the market is heading. The question is not whether open weights will matter. It is whether US-based open-weights models can sustain the economics against Chinese competitors operating under fundamentally different cost structures.
Inkling-Small suggests the answer is: at this performance tier, yes — if you are willing to compete on developer experience, fine-tuning infrastructure, and ecosystem integration rather than raw price alone. The model is not the cheapest option. It is the most capable open-weights model with a complete US-based development stack, and for enterprises that care about provenance, supply chain, and regulatory alignment, that distinction may be worth the premium.