Thinking Machines Lab launches Inkling open-weights model

Thinking Machines made an unusual choice at launch: they published Inkling with an explicit caveat that it is not the strongest model available, closed or open. This is not false modesty - it is a deliberate repositioning of what the product is actually selling. The Tinker platform, launched eight months before Inkling, is the revenue engine. Inkling is the fuel. Every enterprise that fine-tunes Inkling on Tinker generates training data, feedback loops, and platform lock-in that a hosted API relationship with OpenAI or Anthropic cannot replicate. ^[1]

The Bridgewater case study is the clearest expression of this thesis in numbers. Bridgewater's fine-tuned Inkling variant reached 84.7% accuracy on financial reasoning tasks - a domain-specific benchmark that a general-purpose model cannot optimize for centrally - at approximately one-fourteenth the cost of the leading proprietary alternatives. ^[2]The implication is structural: the gap between a fine-tuned open model and a generic closed model widens as domain specificity increases. Thinking Machines is betting that most high-value enterprise AI work is domain-specific, and that the company able to provide the best fine-tuning infrastructure around a strong open base will capture more enterprise value than any company chasing the top of the general intelligence leaderboard.

The launch announcement drew immediate high-engagement response. Thinking Machines' official post and Mira Murati's personal announcement together drew tens of thousands of likes and hundreds of replies in the first hours. On YouTube, early coverage skewed toward hands-on evaluation and architecture walkthroughs, with the Hugging Face official channel and independent reviewers both noting the model's unusual positioning as a fine-tuning foundation rather than a flagship benchmark entry.

Inkling's efficiency advantage over comparable open-weights models is not accidental - it is the product of specific architectural choices. The 66-layer decoder-only transformer uses sparse Mixture-of-Experts with 256 routed experts and 2 shared experts per layer, but only 6 routed experts activate per token. This means the full 975B parameter count is a theoretical capacity; actual inference compute tracks the 41B active parameters, which is why Inkling uses roughly one-third the tokens of Nvidia's Nemotron 3 Ultra for equivalent performance on Terminal Bench 2.1. ^[3]

The controllable reasoning effort system is a less-discussed but significant architectural feature. Developers can pass a reasoning_effort argument ranging from none to max, effectively choosing the compute-per-token budget at inference time. This is not available in most competing open-weights models and has direct cost implications for enterprise deployments: a document classification task can run at minimal effort, while a complex financial analysis runs at high or xhigh. Post-training reinforced this capability through an asynchronous RL regime of over 30 million rollouts using a hybrid Muon/Adam optimizer. ^[4]The native multimodal pretraining from scratch on text, images, audio, and video - rather than patching vision onto a text-only base - means audio and image representations share the same token space and attention layers, which is what enables the nuanced interaction modeling Murati described.

The benchmark results reflect these architecture choices in practice. Inkling scored 97.1% on AIME 2026, 87.2% on GPQA Diamond, 77.6% on SWE-bench Verified, 91.4% on VoiceBench, and 78.0% on the FORTRESS adversarial safety evaluation - the highest among compared open-weights models on that last metric. ^[4]These scores place Inkling ahead of Nemotron 3 Ultra on most dimensions while requiring substantially less inference compute per task.

The competitive subtext of Inkling's launch is geopolitical as much as commercial. Chinese labs - DeepSeek and Kimi in particular - have dominated the open-weights leaderboards for much of 2025 and 2026, producing models with strong benchmark performance under Apache-style licenses that Western enterprises can self-host. Inkling enters as the first US open-weights model to reach rank 41 on the Artificial Analysis Intelligence Index, surpassing Nvidia's Nemotron 3 Ultra (38) and Google's Gemma 4 31B (29). ^[3]That ranking specifically captures US-origin models, positioning Inkling as the domestic alternative for enterprises with data sovereignty requirements or US-sourcing preferences.

The Apache 2.0 licensing decision is central to this competitive framing. Apache 2.0 permits commercial use, modification, and redistribution without royalty, which is the same permissive baseline that made DeepSeek's releases so disruptive. Inkling's post-training also incorporated synthetic data from Kimi K2.5, a detail that illustrates how quickly capabilities now propagate across the open-weights ecosystem regardless of national origin. ^[5]The inference partnerships with TogetherAI, Fireworks, Modal, Databricks, and Baseten at launch signal an attempt to build the same kind of accessible ecosystem that has given Chinese open-weights models rapid enterprise adoption - without requiring organizations to manage the 2TB+ VRAM (BF16) or 600GB VRAM (NVFP4) infrastructure needed for self-hosting at full scale.

Thinking Machines' safety evaluation reached a conclusion that is accurate but structurally uncomfortable: Inkling 'did not present risk of material uplift beyond what's already available in the open-weight ecosystem.' This is a meaningful safety statement, but it measures risk relative to an existing baseline that is itself not risk-free. The open-weights ecosystem already contains capable models that can be fine-tuned for harmful purposes; clearing the bar of 'not worse than what exists' does not address the cumulative risk that each new high-capability open-weights release adds to that baseline. ^[5]

The deeper structural issue is that Apache 2.0 licensing and residual risks from role-play and indirect prompting create a combination that proprietary model providers can avoid: once Inkling's weights are public, Thinking Machines has no centralized patch channel. A vulnerability discovered in GPT-4o or Claude can be mitigated server-side within hours. A vulnerability in Inkling's weights requires either a new model release or reliance on downstream deployers to implement guardrails - neither of which is guaranteed. Lilian Weng's presence as co-founder suggests this tradeoff was made with eyes open, and the model card's explicit acknowledgment of role-play and indirect prompting risks reflects genuine transparency. ^[5]The tension between maximizing adoption through open licensing and maintaining meaningful safety governance over deployed instances is unresolved, and will become more acute as fine-tuned Inkling variants proliferate across enterprise deployments where Thinking Machines has no visibility.