Inside Thinking Machines Inkling: Mira Murati's 975B Open-Weight MoE Foundation Model

In mid-2026, the artificial intelligence research ecosystem experienced a defining release: Thinking Machines Lab, founded by former OpenAI Chief Technology Officer Mira Murati, unveiled its inaugural foundation model—Inkling. Released with fully open weights on Hugging Face, Databricks, and Modal, Inkling is architected as a 975-billion-parameter sparse Mixture-of-Experts (MoE) network that activates only 41 billion parameters per query.

Rather than creating another closed, proprietary API walled garden, Thinking Machines designed Inkling to give enterprise software teams complete sovereign control over their weights, fine-tuning datasets, and inference infrastructure. Paired with its dedicated adaptation framework Tinker and a companion lightweight tier (Inkling-Small), the model establishes a formidable Western open-weights alternative to both closed frontier giants and Chinese open releases like DeepSeek and Qwen.

Here is an architectural deep dive into Inkling's MoE routing, its post-training alignment methodology, and what this means for enterprise AI deployments.

1. The Sovereign Open-Weights Thesis

For two years, enterprise AI adoption has been constrained by a stark dilemma: rely on closed API endpoints with opaque prompt logging and data residency concerns, or deploy smaller open-source models that lagged behind frontier capabilities on complex reasoning tasks.

Inkling bridges this divide through scale-efficient open sparsity:

2. Technical Anatomy: 975B-A41B Sparse Routing

Inkling combines dense representation layers with sparse feed-forward expert routing to balance parametric capacity with inference efficiency.

A. Sparse Gating Formulation

For each token representation $h_t$ at layer $l$, the router computes gating weights over $E$ total expert networks:

With $k=2$, each token routes through exactly two 20.5B experts, capping active forward-pass compute at 41B parameters. This sparsity enables Inkling to run at high throughput across standard 8-GPU nodes using FP8 and NVFP4 quantization.

B. Long-Context Stability via Grouped-Query Attention (GQA)

To prevent KV cache explosion across its 256,000-token context window, Inkling employs 8:1 Grouped-Query Attention:

Reducing KV head count by $8\times$ slashes memory footprint during multi-turn agent sessions from 120 GB down to 15 GB, leaving ample on-chip SRAM for batching parallel requests.

3. Benchmark Comparisons: Inkling vs. Open and Closed Peers

In cross-domain evaluations spanning automated code generation, complex scientific reasoning, and instruction following:

4. The Tinker Adaptation Framework

Alongside model weights, Thinking Machines released Tinker—an open-source framework specifically engineered to customize large-scale MoE models without full-parameter re-training:

1. Expert Specialization Freezing: Identifies and freezes general reasoning experts while updating only domain-specific expert feed-forward layers.
2. Deterministic RLVR Alignment: Employs Reinforcement Learning with Verifiable Rewards to align code generation and structured JSON schemas against compiler checks.
3. Sub-Hour Fine-Tuning: Trains domain adapters on a single 8-GPU node in under 60 minutes.

5. Frequently Asked Questions (FAQ)

What hardware is required to run Inkling 975B?

In FP8 precision, Inkling requires approximately 780 GB of unified GPU memory, fitting comfortably across a standard 8x NVIDIA H100/H200 (80GB/141GB) or Blackwell server node. For single-workstation deployments, the companion Inkling-Small model runs on a single 24GB consumer GPU.

How is Inkling licensed?

Inkling is distributed under a permissive commercial open-weights license, allowing enterprises to self-host, fine-tune, modify, and integrate the model into commercial SaaS applications without royalty fees.

6. Conclusion

Thinking Machines Lab's release of Inkling marks an essential maturation step for enterprise artificial intelligence. By pairing massive 975B MoE parametric capacity with open-weights transparency and the Tinker adaptation engine, Inkling delivers a powerful, sovereign foundation for mission-critical software engineering and autonomous systems.

(Cover Image Courtesy: Unsplash / Deep Neural Network Visualization & AI Architecture)