Thinking Machines’ new model takes the lead in western open-weight AI

Thinking Machines Lab introduced Inkling 2 Small, a 594-billion-parameter model that scored 31 on Artificial Analysis’ Intelligence Index, tied with Gemini 3.8 Flash; the company plans to release its weights next week. Artificial Analysis said that would make it the highest-scoring open-weight model developed outside China under current rankings. Versus its predecessor, automation capabilities rose 41% and coding abilities 20%, while the review noted slower, verbose output and pricing of $0.30 per million input tokens versus a $0.24 class median.

Analysis

The investable signal is not a leaderboard rank; it is whether open weights become good enough to shift enterprise workloads from metered APIs to self-hosted or third-party inference. That substitution can pressure pricing and retention for closed-model providers, including Anthropic (ANT), but total cost of ownership—not token price alone—decides adoption: serving capacity, latency, reliability, security review, and engineering overhead may erase savings. Inkling’s reported slowness and verbosity are meaningful adoption frictions, while the cited improvement in coding and automation warrants monitoring rather than assuming production displacement.

For ANT, this is incremental competitive pressure, not yet evidence of changed revenue or guidance. The broader second-order effect may be more favorable to compute infrastructure if open deployment expands inference demand, but inefficient models can raise cost per useful task and limit that benefit. Next week’s weight release is the near-term catalyst; independent replication, latency/cost-per-completed-task benchmarks, and evidence of customer deployment matter more over the next 1–3 months. Over 6–18 months, repeated quality gains could compress API pricing and increase customer bargaining power. The contrarian point: open-weight capability headlines may overstate commercial substitutability when benchmark scores omit operational costs. Thesis weakens if released weights fail to reproduce reported results or enterprise usage remains limited; it strengthens if deployment data show comparable task success at lower all-in cost and ANT’s pricing or guidance responds.

Key Decisions for Investors

• No immediate directional trade in ANT on this release alone: the article provides no evidence of customer switching, revenue exposure, or a company-specific estimate of pricing sensitivity.
• Set an alert for the weight release next week and independently compare quality, latency, and all-in cost per successfully completed coding/agent task against closed alternatives; do not rely on headline benchmark scores.
• For the next 1–3 months, watch ANT pricing, usage commentary, and guidance for signs of API discounting or slower adoption. A sustained pricing response or customer migration evidence would make a relative underweight in ANT more actionable; absent that, avoid treating competitive capability as realized financial impact.
• Monitor compute suppliers as a second-order watch item, not a trade yet: self-hosted adoption could lift inference demand, but model efficiency and workload utilization are missing to establish a beneficiary.