Inkling: Thinking Machines Ships a 975B-Parameter Model It Admits Isn't the Best

What actually shipped

Thinking Machines Lab — the startup founded by former OpenAI CTO Mira Murati — released its first open-weights model, Inkling, under an Apache-2.0 license. It's a Mixture-of-Experts transformer with 975B total parameters and 41B active per token: each MoE layer routes a token to 6 of 256 experts plus 2 shared experts, so you get frontier-scale capacity without paying frontier-scale inference cost on every forward pass. The model was pretrained on 45 trillion tokens spanning text, images, audio, and video, and supports a context window up to 1M tokens.

Under the hood it's a mix of choices you don't see in every release: interleaved sliding-window and global attention at a 5:1 ratio with 8 KV heads, relative positional embeddings instead of RoPE, and a hybrid Muon/Adam optimizer for training. Weights are available in both the original and NVFP4 checkpoints, and the model already runs on SGLang, vLLM, and llama.cpp, with hosted access through TogetherAI, Fireworks, Modal, Databricks, and Baseten.

The unusually honest pitch

What's notable isn't the parameter count — it's the framing. Thinking Machines states directly that Inkling "is not the strongest overall model available today, open or closed." Instead, the pitch is multimodal capability, controllable thinking effort, and — the real hook — availability on Tinker, their fine-tuning service, at a 50% discount for this model. That's a different sales pitch than most model launches: not "beats GPT-5 on MMLU," but "cheap and easy to specialize into your own agent."

They're also previewing Inkling-Small, a 276B (12B active) variant still in testing, which the lab says matches or exceeds full Inkling on many benchmarks despite the smaller active footprint. Weights aren't out yet — Thinking Machines says they'll ship once evaluation work wraps.

Why this matters if you're building agents

Apache-2.0 is the detail worth underlining: no field-of-use restrictions, no "open-weight but not really open-license" fine print that some competing releases carry. Combined with a 1M-token context window and support across every major open-source inference stack, Inkling is less a chatbot competitor and more raw material — something you self-host, quantize, or fine-tune into a narrow, cheap component of a larger agentic pipeline (a router, a tool-calling layer, a domain-specific coding assistant) without licensing risk or vendor lock-in.

Also worth noting for anyone doing diligence on training data: Inkling's model card and its linked Training Data Documentation are noticeably shorter than what's become standard from US labs, and openly acknowledge the training corpus includes "content that may be subject to intellectual property protection" pulled from the open internet. That's a real provenance disclosure, not a marketing footnote — factor it in if you're evaluating the model for a regulated or IP-sensitive product.

The bigger picture

Inkling lands alongside NVIDIA's Nemotron and Google's Gemma line as part of a push to keep large, genuinely open US-built weights in the conversation against Chinese open-weight models like DeepSeek and Qwen. Whether Inkling becomes anyone's default depends less on benchmarks and more on whether the Tinker fine-tuning story holds up in practice — which is the part builders will actually test.