Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released Inkling on Wednesday, a 975-billion-parameter open-weight AI model that represents the largest American open-source model to date. The model is designed to compete with Chinese large language models (LLMs) like DeepSeek V4 and Kimi K2.6, but accessing its full capabilities requires significant infrastructure investment.

What Makes Inkling Different From Other Open-Weight Models?

Inkling uses a mixture-of-experts (MoE) architecture, meaning it doesn't activate all 975 billion parameters for every task. Instead, each token is routed to just six of 256 specialist experts, plus two shared experts, bringing the active parameter count down to 41 billion per token. This design choice cuts computational costs while maintaining the model's full knowledge base.

The model was trained from scratch using Nvidia GB300 NVL72 systems on 45 trillion tokens of text, images, audio, and video. It supports a one-million-token context window, roughly equivalent to processing 100,000 words at once, which helps it handle large codebases and complex search tasks.

Why Does Running Inkling Still Require a Data Center?

Despite the sparse activation design, the full model checkpoint remains massive. The BF16 (16-bit precision) version occupies approximately 1.9 terabytes, while the NVFP4 quantized version is still 592 gigabytes. Thinking Machines specifies that running the BF16 version requires at least two terabytes of aggregate GPU memory, achievable with eight Nvidia B300 accelerators or sixteen H200 GPUs. The NVFP4 version requires a minimum of 600 gigabytes of VRAM, needing four B300 GPUs or eight H200 GPUs.

Even transferring these files is an infrastructure challenge. At ideal one-gigabit-per-second connection speeds, the BF16 tensors would take approximately 4.2 hours to download, while the NVFP4 version would take about 1.3 hours. A ten-gigabit data-center link cuts those times to roughly 25 and eight minutes respectively, but this remains far from a typical software download experience.

How Can Developers Use Inkling Without Building Their Own Cluster?

Thinking Machines offers three practical pathways for developers who lack data-center infrastructure:

• Tinker Platform: A managed fine-tuning service using LoRA adapters that allows developers to customize the model without retraining the full 975-billion-parameter base. At launch, Tinker offered 64K and 256K context options at discounted rates, with pricing ranging from $1.87 to $11.23 per million tokens depending on the operation and context window.
• Third-Party API Providers: The model is available through Together AI, Fireworks, Modal, Databricks, and Baseten, which handle infrastructure management and offer pay-as-you-go access.
• Self-Hosted Runtime Support: For teams with moderate GPU resources, Inkling supports multiple inference engines including vLLM, SGLang, Miles, TokenSpeed, and Llama.cpp, allowing deployment on smaller clusters than the official minimum.

The trade-off is clear: Tinker lowers the hardware barrier by converting infrastructure into usage charges, but longer-context customization roughly doubles the per-token rates. A team training a fine-tuning job on ten million tokens would pay $56.10 at 64K context or $112.30 at 256K context, though real-world jobs typically involve repeated passes and additional evaluation steps.

How Does Inkling Compare to Proprietary Models and Chinese Competitors?

Thinking Machines claims Inkling is competitive with DeepSeek V4, GLM 5.2, and Kimi K2.6 across various benchmarks, though the company acknowledges that its benchmark charts show it trailing proprietary models like Anthropic's Claude and OpenAI's GPT. The model includes reasoning capabilities, trained using reinforcement learning to use chain-of-thought reasoning before responding. Thinking Machines states it optimized these thinking tokens for efficiency, matching Nvidia's Nemotron 3 Ultra (a 550-billion-parameter open-weight model) on Terminal Bench 2.1 using roughly one-third the tokens.

This efficiency matters for cost. Thinking tokens are billed like any other token, so more efficient reasoning reduces user expenses. However, the longer the model thinks through a problem, the larger the final bill becomes, creating a practical trade-off between quality and cost.

What's the Broader Significance of This Release?

Inkling represents a significant shift in the open-weight AI landscape. Until now, frontier-class open-weight models have been dominated by Chinese companies like DeepSeek. Murati's move from OpenAI to founding Thinking Machines signals a broader trend of AI researchers building independent companies focused on open-source alternatives to proprietary models.

The company is not stopping with Inkling. Thinking Machines is also previewing Inkling-Small, a 276-billion-parameter MoE model with 12 billion active parameters designed for teams prioritizing speed over throughput and quality. The company plans to release its weights once testing is complete.

The model is available for download on Hugging Face and other model repositories under an Apache 2.0 license, making it freely available for commercial and research use. This permissive licensing, combined with full weight transparency and documented architecture, represents a genuine commitment to openness that contrasts with many competitors' more restrictive approaches.