Inkling: Murati’s Open-Weight Bet Lands on Hugging Face

Thinking Machines Lab released Inkling on July 15, 2026 — the first in-house model from the startup Mira Murati founded after leaving OpenAI, and a 975-billion-parameter open-weight bet that the future of enterprise AI is customization, not capability leaderboards.

The specs read like a frontier launch: a Mixture-of-Experts transformer with 41 billion active parameters per token, trained on 45 trillion tokens spanning text, images, audio, and video, released under Apache 2.0 with full weights live on Hugging Face. Multiple outlets describe it as the largest US-built open-weight model released to date. And yet the launch post leads with a disclaimer: Inkling is not the strongest model available, open or closed — and Thinking Machines says so itself.

This guide covers what actually shipped, the customize-don’t-rent thesis behind the deliberate benchmark humility, the architecture and the controllable thinking-effort dial, how the vendor’s comparison chart differs from the independent numbers, what self-hosting a 975B model actually requires, and what a credible US Apache 2.0 entrant changes for teams doing procurement and compliance review.

The launch surface is broad for a first release. Full Inkling weights went live on Hugging Face under Apache 2.0 — full commercial, royalty-free rights to download, modify, and redistribute — alongside an NVFP4 quantized checkpoint targeting NVIDIA Blackwell hardware. The model went live for fine-tuning on Tinker, Thinking Machines’ model-customization platform, the same day. And a set of third-party inference partners — TogetherAI, Fireworks, Modal, Databricks, and Baseten — offered hosted access from day one, with inference-engine support for vLLM, SGLang, TokenSpeed, and llama.cpp.

Inkling arrives as part of a crowded open-weight season — Tencent shipped its Apache 2.0 Hunyuan Hy3 reasoning model earlier this month — but it is the first entrant in that wave from a US lab at this scale.

Most model launches lead with a benchmark chart. Thinking Machines led with a concession. The announcement states, in the company’s own words: “Inkling is not the strongest overall model available today, open or closed. Instead, a combination of qualities makes it a good open-weights base for customization: multimodal capabilities, efficient thinking, and availability on Tinker for fine-tuning.”

That framing is the product strategy. Thinking Machines does not monetize Inkling directly — the revenue vehicle is Tinker, the fine-tuning platform it launched in October 2025. Inkling exists to give Tinker customers a permissively licensed, multimodal, frontier-adjacent base model they can shape into something proprietary. The bet is that enterprises would rather own a customized model than rent a general-purpose one — and that a good-enough open base plus deep customization beats a stronger closed model for many domain-specific workloads.

There is one prior proof point, with a caveat: a joint project with Bridgewater Associates fine-tuned an open model on the firm’s financial expertise via Tinker and scored 84.7% on financial-reasoning tests at roughly one-fourteenth the cost of running a top proprietary model — per the two companies’ own, non-independent evaluation. Treat it as a directional signal for the thesis, not a verified benchmark.

The company behind the bet carries unusual weight for a one-model startup. Thinking Machines was founded in February 2025 by Mira Murati, OpenAI’s former CTO, with co-founders including John Schulman and Barret Zoph, and raised a then-record $2 billion seed round at a $12 billion valuation before shipping any product. It now employs roughly 200 people, having lost two co-founders back to OpenAI in January 2026. TechCrunch reported that a larger raise discussed in late 2025 had stalled by January — the company has declined to discuss its funding picture — which makes Inkling’s reception commercially consequential, not just technically interesting.

Inkling is a 66-layer decoder-only transformer with a sigmoid-routed Mixture-of-Experts design: 256 routed experts plus 2 shared experts, with 6 experts active per token. Attention interleaves sliding-window and global layers at a 5:1 ratio, and the model uses relative positional embeddings rather than RoPE. Per Thinking Machines’ own documentation, the MoE design follows the DeepSeek-V3 sigmoid-router pattern — the same architectural lineage that produced DeepSeek’s V4 series. The model is natively multimodal via encoder-free early fusion: text, images, audio, and video share one training stream rather than bolting a vision encoder onto a text model.

The distinctive interface is controllable thinking effort: developers set a reasoning_effort value from 0.2 to 0.99 to trade reasoning depth against speed and cost, exposed directly in Hugging Face transformers (5.14.0+). Post-training ran over 30 million reinforcement-learning rollouts, during which researchers reported an emergent “chain-of-thought condensation” effect — the model learned to compress its reasoning steps, dropping grammatical filler while keeping accuracy, which cuts latency at a given effort setting.

One candid detail from the launch coverage: early post-training data was bootstrapped using outputs from rival open models — including Moonshot AI’s Kimi K2.5 — before large-scale RL took over. Thinking Machines says its next model will use fully self-contained post-training. It is an honest snapshot of where the open ecosystem is in 2026: even the labs building American open weights are standing on Chinese open weights to get there.

Read the coverage carefully and two distinct benchmark stories emerge. The first comes from Thinking Machines’ own comparison chart, reproduced across the launch press: Inkling beats Nvidia’s Nemotron 3 Ultra — the previous largest US open-weight model — on the benchmarks both were tested on, while trailing the top Chinese open models on coding and reasoning. Press-reported figures show GLM 5.2 ahead on SWEBench Pro and Terminal Bench 2.1, DeepSeek V4 Pro ahead on SWEBench Verified and SimpleQA Verified, and Kimi K2.6 ahead on GPQA Diamond and BrowseComp. The most repeated vendor claim — that Inkling reaches the same Terminal Bench 2.1 coding performance as Nemotron 3 Ultra using roughly a third as many tokens — is Thinking Machines’ own measurement and has not been independently verified.

The second story comes from Artificial Analysis, the independent benchmark aggregator, which scored Inkling 41 on its Intelligence Index — calling it the leading open-weight release from a US lab, three points above Nemotron 3 Ultra (38) and well clear of Gemma 4 31B (29) and gpt-oss-120b (24). The same analysis found Inkling markedly more token-efficient per agentic task than its Chinese rivals, and ahead of Kimi K2.6 and DeepSeek V4 Flash-max on two independently run agentic evals (GDPval-AA v2 and τ³-Banking). Neither story is wrong; they measure different things. The table below keeps the reporting basis explicit for every row.

Two footnotes matter when you cite these numbers. First, Humanity’s Last Exam appears in two variants that are easy to conflate: launch-week reporting puts Inkling at roughly 29.7–30.0% on HLE text-only and 46.0% on HLE with tools — two different test conditions, not a contradiction, and any citation should name the variant. Second, the vendor comparisons above were run at Inkling’s maximum effort setting per the launch materials; independent replication of the full chart did not exist at the time of writing.

That token-efficiency chart is the most decision-relevant independent data point in the launch. For agentic workloads billed per token, a model that spends roughly 25K output tokens per task where rivals spend 37–43K changes the cost calculus even when its raw benchmark scores trail — you are buying outcomes per dollar, not points on a leaderboard. It is also the one place where an independent measurement directly corroborates the vendor’s “efficient thinking” framing rather than merely repeating it.

The headline superlative is well corroborated as far as it goes: multiple outlets independently describe Inkling as the largest American open-weight model released to date, surpassing Nvidia’s Nemotron 3 Ultra (550B total / 55B active, released June 4, 2026), which had held the informal title. But “largest” rests on total parameter count — and Thinking Machines itself is explicit that Inkling is not the most capable model, open or closed. The press-reported coding and reasoning comparisons in the table above put GLM 5.2, Kimi K2.6, and DeepSeek V4 Pro ahead on most raw-capability benchmarks.

The healthy skepticism came from The Register, whose launch analysis noted the model is comparable to the Chinese open-weight leaders in size and capability terms while cautioning readers to take vendor benchmark claims with a grain of salt, since gaming AI benchmarks is not exactly difficult. That is the right posture: after a first half of 2026 defined almost entirely by Chinese open-weight releases, the significance of Inkling is not that it wins — it is that a US lab finally showed up at this scale with an Apache 2.0 license.

Inkling ships with an unusually explicit stance on refusals: it was designed to answer directly on topics “that may be subject to censorship,” and per VentureBeat it was submitted to Cognition’s Propaganda and Censorship Eval, where it showed strong patterns of censorship non-compliance. For teams that have watched open models inherit their training data’s political filtering, that is a differentiator with real procurement weight — and a pointed contrast with some Chinese open-weight alternatives.

The reported safety numbers cut the other way from the anti-censorship framing: 98.6% on StrongREJECT for unambiguous harmful requests, and a 78.0% refusal rate on adversarial FORTRESS queries while maintaining 95.9% compliance on benign look-alike queries — the balance every open model has to strike between refusing attacks and not refusing customers. The model card’s own assessment concluded Inkling “did not present risk of material uplift beyond what’s already available” on CBRN, cyber, and loss-of-control testing, while flagging a residual weakness — an occasional tendency to comply with role-play and indirectly framed prompts on harmful topics — and recommending an external moderation layer such as Llama Guard for production deployments. If you deploy Inkling, that recommendation should be treated as a requirement, not a suggestion.

“Free to download” is not free to run. The full BF16 checkpoint needs more than 2TB of aggregated VRAM — on current NVIDIA hardware, that is a cluster of 8× B300 or 16× H200 accelerators before you serve a single request. The NVFP4 quantized checkpoint cuts the footprint to roughly 600GB (4× B300 or 8× H200), which moves self-hosting from “dedicated AI infrastructure team” territory to “serious but attainable” for a well-funded platform group. The model was trained entirely on Nvidia GB300 NVL72 systems — the launch doubles as a proof point for the gigawatt-scale Nvidia compute partnership Thinking Machines announced in March 2026.

For most teams, the realistic on-ramps are the middle paths. Hosted inference through TogetherAI, Fireworks, Modal, Databricks, or Baseten gets you metered access to the open model without the cluster. Fine-tuning through Tinker gets you a customized variant without owning training infrastructure — though note that only a limited-time launch discount has been documented publicly, so get current list pricing in writing before you model unit economics on it. Independent developer Simon Willison, testing Inkling via the Tinker API, called it “good to see the US open weights ecosystem gain a new viable contender to join NVIDIA Nemotron and Gemma 4.”

The break-even logic is workload arithmetic, not ideology. Metered frontier APIs win at low and spiky volume; self-hosted open weights can win at sustained high volume, where utilization stays high enough to amortize the hardware — and Inkling’s independently measured token efficiency shifts that curve in its favor for agentic workloads. The honest answer for most organizations is to run the comparison on their own traffic profile before committing either way.

The trend underneath this launch is bigger than one model. Through the first half of 2026, open-weight momentum belonged almost entirely to Chinese labs, which left every US and European compliance team with an awkward choice: accept Chinese-origin weights, pay frontier API prices, or settle for much smaller US-built options. Inkling collapses that trilemma for a meaningful class of workloads — the second-source case for open weights no longer requires a geopolitical asterisk. That matters even for teams that never deploy Inkling, because credible alternatives discipline pricing and licensing behavior across every vendor conversation.

Looking forward, the launch sets up a test the whole industry should want answered: can a deliberately mid-benchmark open base plus a customization platform out-compete stronger general models on real enterprise workloads? If Tinker’s customers replicate the Bridgewater pattern across industries, the customize-don’t-rent thesis stops being a slogan and becomes a procurement category — and the pressure on closed vendors shifts from capability to price. If they don’t, Inkling risks being remembered as a very large, very well-licensed model that arrived trailing its rivals. Either way, teams evaluating whether a fine-tuned open model can replace a rented frontier one are making exactly the kind of build-versus-rent decision our AI transformation engagements are structured around: benchmark on your own tasks, price both paths at your real volume, and let the workload decide.