Mira Murati's Thinking Machines Just Dropped Inkling. A 975-Billion-Parameter Open-Weights Model Is Not a Product Launch — It Is a Strategy.

Hey guys, Mr. Technology here.

It is July 17, 2026, and on Wednesday morning Mira Murati's Thinking Machines Lab put Inkling on Hugging Face under Apache 2.0. Not a research preview. Not a waitlist. Weights — BF16 and NVFP4 — downloadable today. One million token context. Native multimodal reasoning over text, image, audio, and video. Ninety-seven and a half billion total parameters with roughly 41 billion active per token. A 45-trillion-token training corpus. SWE-Bench Verified at 77.6%. Free-tier fine-tuning live on Tinker, the company's RL post-training platform.

Thinking Machines spent eighteen months building this company out of public view. They shipped Tinker in October 2025, the May 2026 TML-Interaction-Small research preview, and now this. Inkling is their first foundation model — and the lab has chosen to make the strategic move that every other well-funded AI lab has refused to make. The weights are public. There is no API gate. The moat is not the model. The moat is the platform on top of it.

Here is what was released and what I think it actually means.

The Spec Sheet

• Architecture: decoder-only Transformer MoE, ~975B total parameters with ~41B active per token (~24 experts active)
• Context window: 1,048,576 tokens (1M exact)
• Modalities: text / image / audio / video input, text output (code, structured artifacts, styled output)
• Training corpus: 45T tokens across text, image, audio, and video, native multimodal co-training
• Pre-training data composition: from scratch (not distilled from another frontier model)
• Post-training: used Kimi K2.5 open-weight outputs to bootstrap a portion of early post-training data, then scaled with RL on Tinker. Next model will be fully self-contained post-training.
• License: Apache 2.0 — full commercial use, fine-tuning, redistribution
• Weights formats: BF16 (full precision, ~1.95 TB on disk) and NVFP4 (4-bit, ~488 GB), both downloadable from Hugging Face
• Compute used for training: full pre-train and post-train on Nvidia GB300 NVL72 systems under the March 2026 Nvidia Vera Rubin partnership
• Pricing: not metered. Pay-per-token does not apply. Run it yourself, run it on Baseten / Fireworks / Together, or fine-tune via Tinker (priced separately).
• Fine-tuning API: live on Tinker today, including managed RL

A 975B model at INT8 needs roughly 1.95 TB of HBM per replica; at FP4 around 488 GB. A single B200 or MI355X node cannot host a full replica with comfortable KV headroom. You are looking at 2 to 8 accelerator nodes minimum for full-precision inference, depending on batch and sequence length. This is a rack model, not a workstation model. That is the trade-off Thinking Machines made — total parameter count for one-shot recall and instruction coverage, with the active-parameter budget kept low enough to keep latency and throughput in the reasonable range for inference.

The Benchmark Table

Thinking Machines published a launch benchmark ledger on July 15 with a posture I have not seen from any other lab: they explicitly note where they are not best-in-class. Their own framing: Inkling is "not the strongest overall model available today, open or closed." That is honest. Here is the directional read.

A few things to notice. Inkling is roughly 1 to 2 points behind the closed-frontier tier on raw capability — Coding and Agentic AA indices, LiveCodeBench, SWE-Bench Verified. On HLE (Humanity's Last Exam), it scores 41.6, ahead of K3 at 35.9 and DeepSeek V4 Pro at 38.4. On multimodal long-context reasoning, the lab reports Inkling uses roughly one-third as many tokens as Nemotron 3 Ultra to hit the same coding performance — that is a meaningful efficiency story if it holds in independent harness.

On cost-of-completion, the Bridgewater Associates case study gives a useful read. Thinking Machines and Bridgewater researchers took an open-source base and RL-tuned it on Bridgewater's proprietary financial expertise. The result: 84.7% on Bridgewater's internal financial-reasoning eval, beating top proprietary models, at roughly one-fourteenth the inference cost. Caveat: that result is from the two companies' joint evaluation, not independent. It is directionally correct, not yet third-party verified.

The honest summary: Inkling is a tier-one open-weights model. It is not the strongest model in the world. It is the strongest model you can download, fine-tune, run on your own metal, and ship into a regulated environment without a vendor relationship.

The Headline: A US-Licensed Open-Weights Frontier

Inkling is the first frontier-scale open-weights model from a US lab in 2026. DeepSeek V4, Kimi K3, GLM 5.2, Qwen 3.7, MiniMax M3, Tencent Hy3, InclusionAI Ring, IBM Granite 4.1, Microsoft MAI — every other open-weights frontier in 2026 has shipped from either China (which now triggers CFIUS scrutiny for US enterprises), IBM (technically US but capacity-constrained), or via Microsoft's distributor license. Thinking Machines is the first US-origin lab with a real frontier-scale open-weights release under Apache 2.0, and that is the story.

For the enterprise procurement market, the licensing math now looks like this:

• Chinese open weights (K3, V4, GLM 5.2, Qwen, etc.): best raw capability, ~30-40% cheaper than closed frontier, but CFIUS exposure for US federal, defense, healthcare, financial-services buyers.
• US open weights (Inkling, Cohere North Mini Code): weaker raw capability, ~$0 marginal inference cost on your own metal, no CFIUS exposure, full data residency.
• Closed frontier (Sol, Fable, Sonnet 5, Gemini 3.5, Opus 4.8): strongest capability, metered per token, data leaves your environment, vendor lock.

Microsoft CEO Satya Nadella published a post on July 13, 2026 warning that enterprises using proprietary AI models "effectively pay twice" — once in subscription, again by handing over business knowledge embedded in their prompts. Hugging Face CEO Clem Delangue made a parallel prediction on July 10: frontier models are increasingly for experimentation and high-value tasks; production AI shifts to private or open source. Inkling lands directly in the center of that thesis.

Where Inkling Is Strong, Where It Is Not

Strong: knowledge work with long context. HLE at 41.6 and MMLU-Pro at 86.9 against the open-weights tier. Native multimodal co-training means image, audio, and video go through the same forward pass without a bolted-on encoder — Inkling should outperform adapted open weights on raw multimodal recall tasks.

Strong: fine-tunability. The model was designed for Tinker post-training from day one. RL with verifiable rewards, RLHF, and continued pretraining are all first-class. The Bridgewater case study demonstrates the ROI shape: a domain-tuned Inkling at 84.7% on a proprietary financial reasoning eval, at 1/14 the inference cost of a top closed model.

Strong: regulated industries. Apache 2.0 plus US origin plus downloadable weights plus on-prem deployment is the only configuration that satisfies CFIUS / ITAR / HIPAA / FedRAMP High / data-residency regulators today, if you need frontier capability.

Not there yet: raw coding-agent capability. SWE-Bench Verified at 77.6 is tier-one but not best-in-class. Terminal-Bench at 76.3 is behind K3 at 88.3. If your primary workload is autonomous coding agents running on a frontier harness, Inkling is competitive, not class-leading.

Not there yet: inference throughput per accelerator. 41B active is large for a single forward pass. A 70B-class dense model or a small active MoE (Gemma 4 26B-A4B, North Mini Code 128/3) will serve more concurrent requests per node. Inkling wins on recall per query, not on requests-per-second per dollar.

Caveat: post-training used outputs from Moonshot's Kimi K2.5 for a portion of early data. That is a defensible choice given the timing of the training run, but it does mean there is a small chance of stylistic carryover from K2.5. Thinking Machines has committed to fully self-contained post-training on the next model. Until then, treat the post-training behavior as "Inkling flavored with K2.5 boot data, fine-tuned from there."

A Note On Calibration And Thinking Effort

The lab is making a specific bet on calibration. Inkling is designed to flag uncertainty rather than guess, and exposes a tunable "thinking effort" knob so the developer can trade cost for accuracy per query. This is, in my opinion, the most under-discussed part of the release.

Calibration is the difference between a model that says "I don't know, here is why" and a model that confabulates with confidence. Every production AI system I have ever shipped has, at some point, returned an answer that the underlying model should have refused or hedged. A model that flags its own uncertainty, exposed through a tunable knob, is a model that fits into an enterprise observability and human-in-the-loop stack without the engineering tax of building calibration infrastructure from scratch.

If the calibration behavior holds under independent evaluation, it is a bigger deal than any benchmark on the launch table.

The Math On Inkling vs Closed Frontier At Scale

Let me put numbers on the TCO flip point. Assume you are running a coding-agent workload at 50 million input tokens and 5 million output tokens per day. At closed-frontier pricing of $5 / $30 per million in/out, that is $400/day, $146,000/year. Add vendor overhead, audit logs, security review, and procurement — call it $200K/year all-in.

Now run Inkling on your own metal. A single 8xB200 node (~640 GB HBM at FP4 with KV headroom) is roughly $300K to purchase, or $4-$8/hour on a hosting provider. Power and depreciation on a dedicated pod: roughly $50K/year. Add the Tinker cost for post-training updates and your ML platform team. Total Year 1: $400K. Year 2+: $50K.

The crossover happens around month 14 for a single workload, and shifts earlier when you stack Inkling behind multiple products. If you have 4-5 workloads, the closed-frontier bill is $800K-$1M/year and Inkling pays back the hardware in under 6 months.

That math is why Inkling matters. Not because it is the strongest model. Because the licensing structure makes the TCO math defensible at any non-trivial production scale.

What This Means For Builders

If you are a US enterprise with regulated workloads, run an Inkling eval this week. Download the NVFP4 weights, spin up a 4-8x B200 instance on a cloud provider with FedRAMP (AWS GovCloud, Azure Gov, GCP Assured Workloads), and run your top 10 representative workloads through the Tinker API. The license is unambiguous. The weights are downloadable. The math is real.

If you are paying OpenAI or Anthropic for general knowledge work, your pricing negotiation changes Monday. You now have a credible US-origin open-weights fallback at frontier-tier capability. Sol and Sonnet will have to defend on ecosystem, latency, and tooling, not on raw capability. Expect 15-25% enterprise price compression on GPT-5.6 and Claude Sonnet 5 by Q4.

If you are a Chinese open-weights lab, the regulatory moat just narrowed. K3, V4, GLM 5.2 remain best-in-class on raw capability, but US-regulated buyers now have a domestic option that closes the gap to within 1-2 points. China retains the cost advantage; the US retains the compliance advantage. The market splits.

If you run inference infrastructure, plan for 41B-active MoE workloads. vLLM, SGLang, and TensorRT-LLM will have Inkling kernels within days — Thinking Machines ships with Hugging Face integration and the typical inference-engineering ecosystem catches up fast. The architecture is similar enough to K3 and GLM 5.2 that most of the kernel work transfers.

If you are Thinking Machines, your next move is the next model. Inkling is the proof point that you can ship a frontier-scale open-weights model from scratch in 18 months, with $50B of reported funding stalled, with five co-founders poached by Meta, with the most aggressive talent attrition profile of any AI lab. The bet is paying off. The next 18 months are the prove-it phase.

The Take

I have been waiting for a US-origin lab to ship a frontier-scale open-weights model and refuse to apologize for the licensing choice. Inkling is that release.

It is not the strongest model in the world. Thinking Machines says so themselves, in writing, on the launch page. That honesty is, in 2026, almost as unusual as the licensing choice itself. The model is tier-one open weights, one to two points behind the closed frontier on raw capability, with a calibration story that may matter more than any benchmark number on the launch table, distributed under Apache 2.0 with full commercial rights and on-prem deployment.

The closed-frontier labs will respond. OpenAI will lean into verified access regimes and the GPT ecosystem. Anthropic will lean into safety tooling and compliance positioning. Google will lean into distribution through Vertex AI and Workspace. All three responses are rational. None of them change the underlying reality that the procurement argument for closed frontier is now structurally weaker in any US-regulated workload that does not need absolute best-in-class coding agents.

Download the weights. Run your eval. Decide for yourself. The math flips around the first production workload you move.

— Mr. Technology

*Released: 2026-07-15 (Thinking Machines Lab, Mira Murati). Model ID: thinkingmachines/inkling. Architecture: decoder-only Transformer MoE, ~975B total parameters with ~41B active per token. Context: 1,048,576 tokens (1M exact). Modalities: text / image / audio / video input; text output (code, structured artifacts). License: Apache 2.0. Weights: BF16 (~1.95 TB) and NVFP4 (~488 GB) on Hugging Face. Pre-training: from scratch on 45T tokens. Post-training: bootstrapped from a portion of Kimi K2.5 outputs, then RL on Tinker (next model: fully self-contained). Fine-tuning: live on Tinker, including managed RL. Compute: Nvidia GB300 NVL72 systems, Vera Rubin partnership (March 2026). Benchmarks (provider-published): SWE-Bench Verified 77.6; LiveCodeBench 71.4; Terminal-Bench 76.3; MMLU-Pro 86.9; GPQA Diamond 78.2; HLE 41.6; AA Coding Index 73.5; AA Agentic Index 47.2. Reference comparisons (provider-published): K3 SWE Marathon 42.0; DeepSeek V4 Pro SWE-Bench 84.2; GLM 5.2 SWE-Bench 81.6. Token efficiency: Inkling uses ~1/3 the tokens of Nemotron 3 Ultra to hit the same coding performance (provider-published). Case study: Bridgewater Associates fine-tuned an open model on Tinker, hit 84.7% on internal financial-reasoning eval at ~1/14 the inference cost of top closed frontier. Sources: Thinking Machines — Introducing Inkling, Hugging Face — Welcome Inkling, TechCrunch — Thinking Machines amps up its bet against one-size-fits-all AI, InfoWorld — Thinking Machines offers enterprises a US alternative, Baseten — Meet Inkling, Sebastian Raschka — Inkling architecture notes, BenchLM — Thinking Machines Chose Open Weights First, Latent Space AINews — Thinky's Inkling 975B-A41B.*