The New Enterprise Battle Over Who Owns the Model

Cursor’s All Hands Foreshadowed Model Ambitions
- Cursor’s all-hands revealed plans to become a top-tier model developer and potentially produce a state-of-the-art model by year-end.
- The SpaceX partnership and Grok 4.5 release show Cursor moving from coding-only to broader model ambitions. (Time 0:00:52)
Microsoft Selling Models As Cost And Security Advantages
- Microsoft is training sales to sell MAI models versus OpenAI/Anthropic by emphasizing cost, efficiency, and security integration.
- This aligns with Microsoft’s push to upsell customer-specific fine-tuning (Frontier Tuning) within its ecosystem. (Time 0:07:42)
Inkling Is Built For Customization Not Benchmark Domination
- Thinking Machines Lab released Inkling, an open-weight multimodal model with 975B parameters and a 1M-token context designed for fine-tuning on Tinker.
- TML emphasizes customization and on-prem sovereignty as its primary value versus chasing raw frontier benchmark ranks. (Time 0:12:52)
Pretrained From Scratch With Token Efficiency Advantages
- Inkling was pre-trained from scratch with a small synthetic bootstrap and scores competitively vs open models but trails top closed models on many benchmarks.
- It excels in token efficiency and multimodal (audio/vision) tasks, making it attractive for specific workloads. (Time 0:14:00)
Forward Deployed Fine Tuning As A New Services Niche
- TML’s go-to-market targets companies worried about leaking alpha by offering open weights plus paid fine-tuning services (forward-deployed experts).
- This creates a niche service category: on-prem or bespoke tuning where inference costs and ownership matter. (Time 0:20:05)
Experiment With Open Weights For Data Sovereignty
- Enterprises should consider open-weight models when they need data sovereignty and control over learning, not just peak performance.
- TML markets Inkling plus Tinker as a way to fine-tune privately so companies keep learning exclusive to them. (Time 0:20:16)
The Hidden Costs Make Fine Tuning Risky
- Fine-tuning is harder and costlier than many expect: ongoing data curation, pipeline maintenance, capability regressions, and continuous ops are required.
- The ‘bitter lesson’ suggests scale and generalist pretraining can outcompete bespoke tuning over time. (Time 0:23:33)
Several Forces Are Fueling An Open Weights Moment
- Multiple factors are driving an open-weight moment: viable near-state performance, many funded teams, sovereignty concerns, and geopolitics.
- This convergence increases choice for enterprises across models, harnesses, and architectures beyond a few closed labs. (Time 0:25:38)
Start Experimenting Today To Preserve Enterprise Optionality
- Enterprise buyers should actively experiment with open models and tuning approaches rather than waiting to pick a single vendor.
- Testing multiple harnesses and fine-tuning paths now preserves optionality during this experimental period. (Time 0:27:58)