Thinking Machines Lab has released Inkling, a 975-billion-parameter open weights AI model that instantly becomes the largest American-built open model available to developers. The release marks the first major product from the startup founded by former OpenAI CTO Mira Murati in early 2025.
Inkling ships under an Apache 2.0 license, meaning developers can download, run, fine-tune, and modify it without restrictive terms. That positions the model as a direct alternative to Chinese open weights offerings from DeepSeek, Zhipu, and Moonshot — a gap that no US lab had meaningfully filled until now.
Hardware Requirements and Quantization
Running Inkling at its native 16-bit precision demands more than two terabytes of GPU memory — roughly eight Nvidia B300 accelerators or sixteen H200s. For organizations without that level of hardware, Thinking Machines also released an NVFP4 quantized version that halves the GPU footprint while preserving usable quality.
The model is available for download on Hugging Face and supports a broad set of inference engines at launch, including vLLM, SGLang, Miles, TokenSpeed, and Llama.cpp. Developers can also access it through Thinking Machines’ own Tinker platform, which provides customization and fine-tuning tools. Third-party API providers — TogetherAI, Fireworks, Modal, Databricks, and Baseten — are integrating the model as well.
Architecture and Training
Inkling uses a mixture-of-experts (MoE) architecture that Thinking Machines acknowledges was inspired by DeepSeek-V3. The company trained the model from scratch on Nvidia GB300 NVL72 systems using 45 trillion tokens of text, images, audio, and video.
The architecture routes each token through 256 expert pathways plus two shared experts, with six experts activated per generated token. That yields roughly 41 billion active parameters per token — a design that keeps inference throughput competitive with DeepSeek V4 on equivalent hardware despite Inkling’s larger total parameter count.
The model supports a one-million-token context window, making it suitable for large codebase analysis, long-document reasoning, and retrieval-heavy workloads where short-term memory depth matters.
Benchmarks and Reasoning Efficiency
Thinking Machines claims Inkling is competitive with Chinese frontier open models like DeepSeek V4, GLM 5.2, and Kimi K2.6 across a range of workloads, though the company’s own benchmark charts show it still trails proprietary systems from Anthropic and OpenAI. Benchmark gaming remains a known issue across the industry, so independent evaluation will be the real test.
One notable efficiency claim: Thinking Machines says it tuned Inkling’s reasoning tokens — the chain-of-thought steps the model generates before producing an answer — to match Nvidia’s Nemotron 3 Ultra on Terminal Bench 2.1 while using roughly one-third the tokens. Nemotron 3 Ultra, at 550 billion parameters, had been the largest American open weights model prior to Inkling’s release.
Reasoning tokens improve accuracy and reduce hallucination, but they also inflate inference costs because providers bill for them like any other generated output. Efficient token use directly translates to lower operating costs for anyone running the model in production.
Self-Directed Fine-Tuning
Thinking Machines emphasizes adaptability as Inkling’s core selling point. The model is designed for developers building AI applications, and the company claims it can write its own fine-tuning scripts — refining its behavior, teaching itself new skills, and evaluating its own performance without extensive human intervention.
The Tinker platform provides the tooling layer for this customization, and the Apache 2.0 license removes legal barriers to derivative work. That combination gives enterprises a path to specialized models without depending on a closed API provider’s roadmap or pricing changes.
A Smaller Variant Is Coming
Alongside Inkling, Thinking Machines is previewing Inkling-Small — a 276-billion-parameter MoE model with 12 billion active parameters. The smaller variant targets deployments where latency matters more than peak quality. Weights for Inkling-Small will be released once final testing completes.
The company describes Inkling as the first of several models in development, signaling that Murati’s lab intends to build a full product lineup rather than a single flagship release.
What Happens Next
Inkling’s arrival reshapes the open weights landscape in two ways. First, it gives American enterprises a domestic alternative to Chinese models for sovereignty-sensitive workloads — relevant as lawmakers push tighter curbs on Chinese chipmakers and as regulators scrutinize foreign AI supply chains. Second, it applies competitive pressure on OpenAI and Anthropic, whose closed models still lead on benchmarks but now face a credible open challenger from a team led by a former OpenAI executive.
Watch for independent benchmark results from the open source community over the coming weeks — particularly on coding and long-context tasks where Inkling’s million-token window and reasoning efficiency could differentiate it. Enterprise adoption will depend less on benchmark scores and more on whether the Tinker platform delivers on its fine-tuning promises at production scale. If Inkling-Small lands with strong latency numbers, it could capture the mid-tier deployment market that current open models struggle to serve efficiently.
— David Kim, technology desk, AXO News