# Inkling (Thinking Machines Lab)

> Source: https://aiwiki.ai/wiki/inkling
> Updated: 2026-07-27
> Categories: AI Models, Large Language Models, Open Source AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution to "AI Wiki (aiwiki.ai)".

**Inkling** is a multimodal large language model released on 2026-07-15 by [Thinking Machines Lab](/wiki/thinking_machines_lab), the startup founded by former OpenAI chief technology officer [Mira Murati](/wiki/mira_murati). It was the company's first production model, following earlier research previews, it was trained from scratch, and it shipped with downloadable weights under the [Apache License 2.0](/wiki/apache_license) [1][2][6]. Inkling is a sparse [mixture-of-experts](/wiki/mixture_of_experts) transformer with 975 billion total parameters and 41 billion active per token, accepts text, image and audio input, and supports a context window of up to 1 million tokens [1][3]. Thinking Machines did not present it as the best model on the market. The announcement says plainly that Inkling "is not the strongest overall model available today, open or closed," and positions it instead as a base model that organizations customize through [Tinker](/wiki/tinker), the company's paid fine-tuning platform [2][8]. The name is shared with two unrelated things in technology: Inkling, a San Francisco company founded in 2009 that sells digital content authoring software and was acquired by Echo360 in 2024 [14], and Inkling, the domain-specific language in which teaching curricula are written for Microsoft's Project Bonsai [15]. Neither has any connection to Thinking Machines Lab.

## Release and availability

The model card is dated 2026-07-15, and coverage from TechCrunch, Fortune and Simon Willison places the launch on that Wednesday [1][8][9][5]. Weights went up on [Hugging Face](/wiki/hugging_face) in two repositories: `thinkingmachines/Inkling`, the original bfloat16 checkpoint, and `thinkingmachines/Inkling-NVFP4`, an [NVFP4](/wiki/nvfp4) checkpoint intended for efficient inference on [NVIDIA Blackwell](/wiki/nvidia_blackwell) hardware [1][3].

Running it locally is not cheap. The model card gives a minimum of roughly 2 TB of aggregate VRAM for the BF16 checkpoint (8x NVIDIA B300 or 16x H200) and about 600 GB for the NVFP4 version (4x B300 or 8x H200) [1]. Local inference is supported through [SGLang](/wiki/sglang), [vLLM](/wiki/vllm), TokenSpeed, Unsloth, llama.cpp and the Hugging Face transformers library [1][2]. Hosted access at launch came from [Together AI](/wiki/together_ai), [Fireworks AI](/wiki/fireworks_ai), [Modal](/wiki/modal), [Databricks](/wiki/databricks) and [Baseten](/wiki/baseten), alongside Thinking Machines' own Tinker API [2].

The company does sell metered inference. Tinker offers Inkling at two context tiers, 64K and 256K tokens, with a 50% introductory discount at launch [2]. [Artificial Analysis](/wiki/artificial_analysis) lists the 64K tier at $1.87 per million input tokens, $0.374 per million cached input tokens and $4.68 per million output tokens; every rate doubles at the 256K tier, to $3.74 input, $0.748 cached and $9.36 output [6][7]. The 1 million token context applies to the open weights rather than the hosted tiers [6].

## Architecture

| Property | Value |
| --- | --- |
| Total parameters | 975B [1] |
| Active parameters per token | 41B [1] |
| Layers | 66, decoder-only [1][4] |
| Routed experts | 256, with 6 active per token [1][4] |
| Shared experts | 2 [1][4] |
| Hidden size | 6144 [4] |
| Vocabulary size | 201,024 [4] |
| Checkpoint precision | bfloat16, plus an NVFP4 variant [2][4] |
| Maximum context | 1,048,576 tokens [3][4] |
| Input modalities | Text, image, audio [1] |
| Output modality | Text only [1] |
| License | Apache 2.0, with a separate Model Acceptable Use Policy [1][16] |

The `config.json` in the Hugging Face repository confirms the expert configuration directly: 256 routed experts, 6 selected per token, 2 shared, 66 hidden layers, hidden size 6144, and a text `model_max_length` of 1,048,576 [4]. Sebastian Raschka's architecture notes, written from the model card and config, describe a hybrid attention stack of 55 local-attention layers with a 512-token window plus 11 global layers, and two choices that set Inkling apart from most 2026 MoE models: a learned, input-dependent relative position bias in place of rotary embeddings, and kernel-4 convolutions applied after the key and value projections and on the attention and MLP outputs [10]. He also notes an extra RMSNorm placed immediately after the token embeddings [10].

At 41B active out of 975B total, Inkling activates about 4.2% of its parameters per token, which makes it somewhat less sparse than [Kimi K2.5](/wiki/kimi_k2_5) at roughly 3.2% (32B active out of 1T), and it carries about 231B more total parameters than [GLM-5.2](/wiki/glm_5_2) for a nearly identical active footprint [10]. Raschka's reading is that with conventional grouped-query attention and a larger active parameter count, raw decoding speed is unlikely to be Inkling's competitive edge [10].

Images are handled by a hierarchical patch encoder built on 40x40 pixel patches; audio is converted to dMel spectrograms and encoded as discrete tokens [1][2]. The model card gives optimal input ranges of 40 to 4096 pixels for images and 16 kHz WAV audio up to roughly 20 minutes [1]. Video appears in the training mix but the published input specification covers text, image and audio [1][2].

## Training

Thinking Machines says Inkling was pretrained on 45 trillion tokens spanning text, images, audio and video [2]. Pretraining used a hybrid optimizer setup, the [Muon optimizer](/wiki/muon_optimizer) for the large weight matrices and Adam for everything else, and ran on [NVIDIA GB300 NVL72](/wiki/nvidia_gb300_nvl72) systems [2][8]. Total training compute has not been disclosed.

Post-training is where the announcement is unusually candid. The initial supervised fine-tuning phase used synthetic data generated by other open-weights models, including Moonshot AI's Kimi K2.5, before a large-scale asynchronous reinforcement learning phase covering more than 30 million rollouts [2]. TechCrunch reported that the company has committed to a fully self-contained post-training pipeline for its next model [8].

The training data description in the model card is thin by design. It says the corpus draws on "publicly available sources, acquired from third-parties, or synthetically generated or augmented" content, without naming datasets, and describes curation only in general terms as "cleaning, processing, and modifying datasets," including deduplication and filtering [1]. The card says nothing at all about copyright or intellectual property in the training set. There is no arXiv technical report; the model card and the launch post are the only first-party technical documents.

## Thinking effort

Inkling exposes a continuous effort parameter that trades output tokens against accuracy. Thinking Machines published a sweep from effort 0.2 to 0.99 and states that all of its headline benchmark numbers were produced at effort 0.99 and temperature 1.0, which is the best case rather than a default [2]. Cost scales with the setting, so the published scores and the published prices do not describe the same operating point.

## Benchmarks

All figures in this table are vendor-reported, taken from the Inkling model card at effort 0.99 [1].

| Benchmark | Inkling | Category |
| --- | --- | --- |
| AIME 2026 | 97.1% | Reasoning [1] |
| [GPQA Diamond](/wiki/gpqa_diamond) | 87.2% | Reasoning [1] |
| Humanity's Last Exam, text only | 29.7% | Reasoning [1] |
| Humanity's Last Exam, with tools | 46.0% | Reasoning [1] |
| [SWE-bench](/wiki/swe_bench) Verified | 77.6% | Agentic coding [1] |
| [SWE-Bench Pro](/wiki/swe_bench_pro) Public | 54.3% | Agentic coding [1] |
| Terminal-Bench 2.1 | 63.8% | Agentic coding [1] |
| GDPval-AA v2 | 1238 Elo | Agentic general [1] |
| MCP Atlas | 74.1% | Agentic general [1] |
| Tau 3 Banking | 23.7% | Agentic general [1] |
| Toolathlon Verified | 45.5% | Agentic general [1] |
| BrowseComp | 77.1% | Agentic general [1] |
| SimpleQA Verified | 43.9% | Factuality [1] |
| MMMU Pro (Standard 10) | 73.5% | Vision [1] |
| CharXiv RQ | 78.1% (82.0% with Python) | Vision [1] |
| MMAU | 77.2% | Audio [1] |
| VoiceBench | 91.4% | Audio [1] |
| IFBench | 79.8% | Instruction following [1] |
| FORTRESS Adversarial | 78.0% | Safety [1] |
| FORTRESS Benign | 95.9% | Safety [1] |
| StrongREJECT | 98.6% | Safety [1] |

The model card's own comparison columns include Nemotron 3 Ultra, Kimi K2.5, Kimi K2.6, GLM-5.2, [DeepSeek V4](/wiki/deepseek_v4) Pro, Gemini 3.1 Pro (high), Claude Fable 5 (max) and GPT-5.6 Sol (max/xhigh), and they do not flatter Inkling everywhere [1]. On the hardest reasoning set it trails GLM-5.2 badly, 29.7% against 40.1% on text-only Humanity's Last Exam, while beating it on instruction following, 79.8% against 73.3% [10]. On vision it sits below [Kimi K2.6](/wiki/kimi_k2_6) (73.5% versus 79.0% on MMMU Pro) and well below the closed frontier models [1]. SimpleQA Verified at 43.9% is a real weakness for factual recall, though it beats Nemotron 3 Ultra's 32.4% [1].

### Independent evaluation

Artificial Analysis scored Inkling at 41 on its Intelligence Index, ranking it #10 of 97 tested configurations and calling it the leading [open-weights](/wiki/open_weights) release from a United States lab [6][7]. For context, AA put [Nemotron 3](/wiki/nemotron_3) Ultra at 38, [Gemma 4](/wiki/gemma_4) 31B at 29 and gpt-oss-120b at 24 [6]. On agentic work AA measured a GDPval-AA v2 Elo of 1238 for Inkling, ahead of Kimi K2.6 at 1190 and DeepSeek V4 Flash max at 1189, and 24% on Tau-3 Banking, ahead of both [6].

Token economy is muddier than the launch coverage suggested. AA's launch article reported Inkling averaging 25,000 output tokens per Intelligence Index task, against 43,000 for GLM-5.2 max, 38,000 for Kimi K2.6 and 37,000 for DeepSeek V4 Pro max [6]. AA's model page for the xhigh configuration, by contrast, records 130 million output tokens across the full index versus a 98 million median, and describes the model as somewhat verbose [7]. The two figures come from the same evaluator at different effort settings, and readers comparing running costs should check which one applies. AA measures output speed at 88.4 tokens per second, #18 of 97 [7].

## Inkling-Small

Alongside the main release Thinking Machines previewed **Inkling-Small**, a 276B-parameter MoE with 12B active, trained with a similar recipe and improved pretraining data and aimed at lower cost and latency [2]. Its weights were not published at launch; the company said they would follow once testing was complete [2][5]. As of 2026-07-27 no separate Inkling-Small repository had appeared in the Thinking Machines organization on Hugging Face [18].

## Business model

Thinking Machines does not treat Inkling as a product to be sold by the token so much as an on-ramp. Fortune reported that the model itself is released freely and revenue comes from Tinker, the fine-tuning platform launched in October 2025 [9][13]. Tinker exposes a low-level training API over open-weight models and fine-tunes them with [LoRA](/wiki/lora), training a small add-on rather than changing all of the original weights [19]. TechCrunch reported the company also intends to take a cut of the hosting ecosystem that grows around the weights [8]. The stated thesis is that models organizations adapt for themselves will beat one-size-fits-all systems [8].

The case study the company points to is Bridgewater Associates, which TechCrunch reported built a financial-reasoning model on Tinker that scored 84.7% on financial reasoning tests, beating proprietary systems at roughly one fourteenth of the running cost [8]. That number comes from the vendor side of the relationship and has not been independently reproduced.

The backdrop is a company that has raised a lot and shipped little. Thinking Machines was founded in February 2025 and raised $2 billion at a $12 billion valuation in July 2025 in a round led by Andreessen Horowitz, with Nvidia, AMD, Cisco, Jane Street and the government of Albania among the investors [9][13]. TechCrunch reported a subsequent round at a reported $50 billion valuation had stalled by January 2026, and that the company had roughly 200 employees at launch [8]. It has also lost people: co-founders Barret Zoph and Luke Metz returned to OpenAI in January 2026, and Andrew Tulloch left for Meta Superintelligence Labs [13]. In March 2026 Nvidia announced a strategic partnership including an undisclosed equity investment and a multi-year agreement covering one gigawatt of Vera Rubin capacity [8][13].

## Reception

Inkling landed in a crowded week. Moonshot AI announced [Kimi K3](/wiki/kimi_k3) the following day, and on 2026-07-17 Artificial Analysis scored that model at 57 on the same index, third place at publication and sixteen points above Inkling [12]. Moonshot said K3's weights would follow by 2026-07-27 [17]; as of that date no K3 repository had appeared in Moonshot's Hugging Face organization. Nathan Lambert's assessment was blunt: Inkling is "strong but not in the same class as Kimi K3," and he used it as evidence that well-funded American labs are still trailing Chinese open-weight releases that raised orders of magnitude less capital [11]. Simon Willison's writeup was more sympathetic to the framing, noting that the lab is explicit about not chasing the top of the leaderboard and is selling a customizable base instead [5].

Two criticisms recur. The first is that the release is open in the weights sense only: there is no training data disclosure, no reproducible pipeline, and no paper, and the Apache 2.0 tag sits next to a separate Model Acceptable Use Policy that restricts uses such as surveillance and automated decisions affecting individual rights [1][16]. Anyone comparing it against other entries in an [open-weight license comparison](/wiki/open_weight_license_comparison) should read both documents. The second is practical: a 975B model needing 600 GB of VRAM in its quantized form is not something most teams will run themselves, so "open weights" in this case mostly means portability between cloud providers rather than local control.

The company's own framing of what it optimized for includes calibrated uncertainty and resistance to censorship, which it groups under "epistemics," alongside safety refusal behavior [2]. The FORTRESS Benign score of 95.9% against Claude Fable 5's 55.1% is the clearest quantitative expression of that preference: Inkling is tuned to refuse less on borderline-but-harmless prompts, at the cost of a lower adversarial refusal rate than the most conservative closed models [1].

## See also

- [Thinking Machines Lab](/wiki/thinking_machines_lab)
- [Tinker (Thinking Machines Lab)](/wiki/tinker)
- [Open weights](/wiki/open_weights)
- [Kimi K3](/wiki/kimi_k3)
- [GLM-5.2](/wiki/glm_5_2)
- [MiniMax M3](/wiki/minimax_m3)
- [Qwen](/wiki/qwen)
- [Multimodal Model](/wiki/multimodal_model)

## References

1. "Inkling Model Card." Thinking Machines Lab, 2026-07-15. https://thinkingmachines.ai/model-card/inkling/
2. "Inkling: Our Open-Weights Model." Thinking Machines Lab, 2026-07-15. https://thinkingmachines.ai/news/introducing-inkling/
3. "thinkingmachines/Inkling." Hugging Face, 2026-07-15. https://huggingface.co/thinkingmachines/Inkling
4. "thinkingmachines/Inkling config.json." Hugging Face, 2026-07-15. https://huggingface.co/thinkingmachines/Inkling/blob/main/config.json
5. Willison, Simon. "Inkling: Our open-weights model." simonwillison.net, 2026-07-16. https://simonwillison.net/2026/Jul/16/inkling/
6. "Thinking Machines has released Inkling, the new leading U.S. open weights model." Artificial Analysis, 2026-07-15. https://artificialanalysis.ai/articles/thinking-machines-has-released-inkling-the-new-leading-u-s-open-weights-model
7. "Inkling: Intelligence, Performance & Price Analysis." Artificial Analysis, accessed 2026-07-27. https://artificialanalysis.ai/models/inkling
8. "Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling." TechCrunch, 2026-07-15. https://techcrunch.com/2026/07/15/thinking-machines-amps-up-its-bet-against-one-size-fits-all-ai-with-its-first-open-model-inkling/
9. "Murati's Thinking Machines releases first AI model for broad use." Fortune, 2026-07-15. https://fortune.com/2026/07/15/what-is-mira-murati-thinking-machines-first-ai-model-inkling/
10. Raschka, Sebastian. "Inkling: A New Open-Weight 975B MoE with a Few Surprises." sebastianraschka.com, 2026-07-16. https://sebastianraschka.com/blog/2026/inkling-architecture-benchmark-notes.html
11. Lambert, Nathan. "Kimi K3: The open-weights escalation." Interconnects, 2026-07. https://www.interconnects.ai/p/kimi-k3-the-open-weights-escalation
12. "Kimi K3 achieves #3 in the Artificial Analysis Intelligence Index, comparable to Opus 4.8 and GPT-5.5." Artificial Analysis, 2026-07-17. https://artificialanalysis.ai/articles/kimi-k3-achieves-3-in-the-artificial-analysis-intelligence-index-comparable-to-opus-4-8-and-gpt-5-5
13. "Thinking Machines Lab." Wikipedia, accessed 2026-07-25. https://en.wikipedia.org/wiki/Thinking_Machines_Lab
14. "Inkling (company)." Wikipedia, accessed 2026-07-27. https://en.wikipedia.org/wiki/Inkling_(company)
15. "Machine teaching with Microsoft's Project Bonsai." InfoWorld. https://www.infoworld.com/article/2266017/machine-teaching-with-project-bonsai.html
16. "Model Acceptable Use Policy." Thinking Machines Lab, accessed 2026-07-27. https://thinkingmachines.ai/model-acceptable-use-policy/
17. "Kimi K3." Moonshot AI technology blog, 2026-07-16. https://www.kimi.com/blog/kimi-k3
18. "Thinking Machines Lab models." Hugging Face organization page, accessed 2026-07-27. https://huggingface.co/thinkingmachines/models
19. "Tinker." Thinking Machines Lab, accessed 2026-07-27. https://thinkingmachines.ai/tinker/