Thinking Machines ships its first model — and it's the availability, not the benchmarks, that matters. Inkling is a base model built for fine-tuning, and it lands right as Microsoft trains its salesforce to attack the frontier labs on data trust. The new enterprise fight isn't which model you rent — it's who owns the one you customize.
The next enterprise battle is over who owns the model you fine-tune.
Thinking Machines released Inkling — a mediocre-on-benchmarks but strategically pointed open-weights model designed to be the base for its Tinker fine-tuning platform. It arrives the same week Microsoft trains its salesforce to argue that companies shouldn't trust OpenAI or Anthropic with their data. Two things converge: enterprises increasingly want cost control and data sovereignty, and multiple labs are racing to sell fine-tuning as the answer. Whether closed (Microsoft Frontier Tuning) or open (Inkling + Tinker), the question is no longer just which frontier model to rent — it's how much independence you want over the model itself. The bitter lesson may still win, but the experimental period is real and there's alpha for nimble teams.
Mira Murati's lab released Inkling — a 975B-parameter MoE (41B active) with a 1M-token window that reasons natively across text, images, and audio — plus a 12B Inkling Small in preview. It's the first in a planned family and was pre-trained from scratch, aside from a small bootstrap distilled from Kimi K2.5.
— Thinking Machines Lab, Inkling launch post. TML positioned Inkling honestly: a broad, balanced foundation model whose value is a combination of qualities — multimodal capabilities, efficient thinking, and availability on Tinker for fine-tuning — that make it a good open-weights base for customization.
Inkling scores 29.7% on Humanity's Last Exam (46% with tools), 54.3% on SWE-Bench Pro, and 1238 on GDPVal — ahead of Nemotron and Kimi K2.5 but well behind GLM 5.2, 5.6 Sol, and Fable 5. Artificial Analysis ranked it 19th at 41, while noting it uses ~two-thirds the tokens per task of GLM 5.2 and DeepSeek V4 Pro.
— Professor Ethan Mollick. Ethan Mollick reported it failed the LEM test that every frontier model has passed since DeepSeek-R1 and Sonnet 3.5, and that its chain of thought went haywire even on simple requests. First impressions were broadly unflattering.
— Jack Morris. Jack Morris argues Kimi, GLM, Qwen, and Nemotron all distill — making Inkling a fully different tech stack and the first pure open frontier coding model. A community note added TML did use a small bootstrap from open models like K2.5, and that Llama was also trained without lab distillation.
NLW's read: as enterprises pay more attention to open-weights models, the fact that Inkling is an American model not primarily distilled from a closed lab becomes meaningful — especially once lawyers and risk teams weigh in on data sovereignty.
— Nathan Lambert. Nathan Lambert frames Inkling as the integration of post-training services across TML's entire stack — "one of the best open model business stories to date." The model was designed to be the base model for the Tinker fine-tuning platform.
— Jeffrey Emanuel. Jeffrey Emanuel argues open weights let companies run models on their own infrastructure without leaking data, and TML's clever answer to open-weights monetization is charging for fine-tuning a company's model on its own internal data — keeping the benefits exclusive to that company.
— Daniel Kaplan. Daniel Kaplan sees a new services category: self-serve Inkling + Tinker for teams with AI researchers, plus a high-priced forward-deployed fine-tuner for those without — scaling to big companies, AI startups protecting their alpha, and app-layer companies chasing cheaper tokens.
NLW notes Microsoft Frontier Tuning on MAI models rides on decades of built-up enterprise trust — an easier leap than trusting OpenAI or Anthropic — but it still looks very different from an open-weights option like Inkling that offers real independence and sovereignty.
— Simon Smith. Simon Smith is unconvinced: fine-tuning is far more effort than people think, ongoing to handle edge cases and model updates, and a big general model plus a bit of context often beats hard-won fine-tunes as token costs collapse. Teams frequently ignore the fully loaded cost of curating data and maintaining pipelines.
NLW cautions that acknowledging real challenges in data sovereignty and token cost doesn't mean Microsoft Frontier Tuning or Tinker is the ultimate solution. The flip side: as an open-weights model, Inkling addresses both the cost and sovereignty sides at once, whereas a closed fine-tuning solution may be caught in between.
— Sriram Krishnan, former White House AI advisor and a16z investor. Sriram Krishnan lists the forces behind open source's moment: you can now catch near-SOTA with clean lineage; many well-funded teams exist; organizations fear enabling a future competitor; open source is a slider you tune; companies now worry about ballooning token costs; and countries want frontier tokens inside controlled environments.
NLW argues the decision now spans models, the harnesses they live in, and complex multi-model architectures including fine-tuning approaches. Not every flower will bloom, and the juggernauts may absorb the best ideas — but the experimental period will shape the eventual solutions, with real alpha for nimble teams.
NLW's bottom line for buyers: no need to rush into Tinker or Microsoft Frontier Tuning or switch everything to OpenRouter, but if you make enterprise buying decisions and aren't at least experimenting with or closely watching these fine-tuning and open-weights options, you're missing real opportunity.
The Information reported on a May all-hands where CEO Michael Truell said Cursor aims to produce a state-of-the-art model by year end and accrue a real compute advantage by 2027 — pushing the frontier itself, not staying limited to coding. SpaceX wanted Cursor for its brand, enterprise relationships, and go-to-market team.
Setting aside recent data-retention controversy, the first SpaceX AI model trained in partnership with Cursor has been well received and is competitive. Cursor is also rumored to be building a competitor to Claude Cowork — its first big expansion beyond coding.
SpaceX traded down to $133 Wednesday, breaking the $135 IPO price before recovering to close at $135.27 — now down 33% from its all-time high, costing Musk his trillionaire status. Insider lockups don't start for another month.
Advisors reportedly told Altman a trillion-dollar market cap is unlikely, nudging OpenAI toward next year. Anthropic has appointed investment banks, added a couple billion in revolving credit, and is meeting investors — still targeting an IPO as early as September or October.
The new 4-billion-parameter open model is built to run on edge devices inside robots, functioning as both a world model and a vision-language model. It arrives alongside a fresh wave of robot designs and NVIDIA's expanded Toyota partnership across robots, smart cities, and factories.
A just-launched company called Chip combines AI and autonomy to turn a golf-cart-style vehicle into a robot — the launch video shows a dad sending it to fetch snacks at a soccer game or park itself after an Uber home. A herald of how AI integrates into the physical world.
Per The Information, Apple is far along in an effort to acquire a chipmaker to build AI server chips, having spoken with bankers and approached semiconductor startups. Its own server-grade Baltra chip has been delayed, and it currently outsources to Google Cloud on NVIDIA hardware — a striking shift for a company whose biggest deal ever was $3B for Beats.
NVIDIA just paid $20B for Groq, Cerebras is valued at $40B, and Broadcom's $1.8T valuation would make it a merger with huge antitrust exposure. Tenstorrent is a possible natural fit given ex-Apple designer Jim Keller — but no solid rumors yet.
Per Bloomberg, Microsoft's new fiscal-year strategy has sales staff pushing MAI models' cost-effectiveness and its vertically integrated stack — telling customers Claude is "slower and less accurate" and lacks proper security integrations in Office, while sidestepping overall model quality.
NLW reads Microsoft's move as aligning with Nadella's argument that the frontier labs have an incentive to build competing services — and as a foot in the door for its own models before upselling customer-specific fine-tuning via Microsoft Frontier Tuning.