The Inkling model card is the formal release document for Thinking Machines Lab's first model, Inkling, published on the company's website on July 15, 2026 (thinkingmachines.ai). It is organized in eight numbered sections (general information, model properties, methods of distribution, training, evaluations, safety, bias, risks and limitations, and legal) and is the primary source for the model's architecture, hardware requirements, evaluation results and safety conclusion. The card is a live web page with no revision date; a capture of 2026-09-19 differs from a condensed capture of 2026-07-16 in three cells and one row of its evaluation table (see Versions of the card, below).
| Field | Value |
|---|---|
| Provider | Thinking Machines Lab, Inc. |
| Released | July 15, 2026 |
| License | Apache 2.0 |
| Model type | Multimodal autoregressive transformer |
| Parameters | 975B total / 41B active |
| Context window | up to 1M tokens |
| Input modalities | Text, images, audio |
| Output modalities | Text only |
| Languages | English, general multilingual support |
Intended uses
The card describes Inkling as a general-purpose multimodal model that accepts text, image and audio inputs and generates text, intended for use in English and other languages and across multiple programming languages. Its stated audience is developers building AI-powered applications, including agentic and tool-use systems, coding assistants, chatbots and retrieval-augmented generation systems, and it is described as suitable for general conversational use, instruction following and other natural-language and multimodal tasks. The card states that the model is released with open weights to support research, fine-tuning and integration into third-party products by downstream developers. Use is governed by a Model Acceptable Use Policy (thinkingmachines.ai).
Architecture
Inkling is described as a 66-layer decoder-only transformer with a sparse Mixture-of-Experts feed-forward backbone: each token is routed to 6 of 256 experts, plus 2 shared experts active on every token. Attention is a hybrid of local and global layers. The model is natively multimodal — images encoded via a hierarchical patch encoder, audio via discrete token encoding — with all modalities projected into a shared hidden space and processed jointly by the decoder. Supported numerics are BF16, MXFP8, and NVFP4. See Mixture of Experts (MoE).
Input constraints are specified: text as UTF-8; images as any pixel-based input, with each dimension between 40px and 4096px for optimal performance; and audio as 16kHz WAV, with length within 20 minutes for optimal performance. Output is UTF-8 text.
Distribution and self-hosting
The model is available through API access on Tinker, Thinking Machines' fine-tuning service, through third-party inference providers, and as downloadable weights on Hugging Face (huggingface.co). For Tinker access the card points to the Tinker Cookbook on GitHub and the tml-renderers Python package; for third-party providers it refers to the providers' own documentation. The card states self-hosting requirements for the two checkpoint formats in which the weights are distributed:
| Checkpoint | Minimum VRAM | Example configuration |
|---|---|---|
| BF16 | ≥2 TB aggregated | 8× NVIDIA B300 or 16× NVIDIA H200 |
| NVFP4 quantized | ≥600 GB | W4A4 on 4× B300 (W4A4 additionally requires SM100+ architecture), or W4A16 on 8× H200 |
Running the model directly on GPU hardware requires an inference deployment framework (SGLang, vLLM, TokenSpeed, Unsloth, or Hugging Face) with its dependency libraries.
Training data
The card describes data types as text, images, audio, and video, with provenance "drawn from publicly available sources, acquired from third parties, or synthetically generated or augmented," where publicly available data "includes content from the public internet and publicly accessible repositories." Curation is described as cleaning, processing, and modifying datasets; the steps vary by data type and may include deduplication and filtering "to remove junk or low-quality data, or to advance safety or other objectives." A separate Training Data Documentation page is referenced in the legal section (thinkingmachines.ai). See Training Data Walls.
Evaluations
All Inkling scores are reported at effort=0.99, against a comparison set of open-weight models (Nemotron 3 Ultra, Kimi K2.5, Kimi K2.6, GLM 5.2, DeepSeek V4 Pro) and closed-weight models (Gemini 3.1 Pro (high), Claude Fable 5 (max), GPT 5.6 Sol (max/xhigh)). The table groups 25 rows under eight headings: Reasoning (HLE text only, HLE with tools, AIME 2026, GPQA Diamond); Agentic coding (SWE-bench Verified, SWE-bench Pro Public, Terminal Bench 2.1 best harness); Agentic general (GDPVal-AA v2, MCP Atlas, Tau 3 Banking, Toolathlon Verified, BrowseComp with context management); Factuality (SimpleQA Verified, AA Omniscience); Chat (IFBench, Global-MMLU-Lite); Vision (MMMU Pro Standard 10, CharXiv RQ, CharXiv RQ with python); Audio (Audio MC, MMAU, VoiceBench); and Safety (FORTRESS Adversarial, FORTRESS Benign, StrongREJECT). Selected rows, as printed on the live page:
| Benchmark | Inkling | GLM 5.2 | DeepSeek V4 Pro | Gemini 3.1 Pro | Fable 5 | GPT 5.6 Sol |
|---|---|---|---|---|---|---|
| HLE (text only) | 29.7% | 40.1% | 35.9% | 44.7% | 53.3% | 47.2% |
| HLE (with tools) | 46.0% | 54.7% | 48.2% | 51.4% | 64.5% | 55.0% |
| AIME 2026 | 97.1% | 99.2% | 96.7% | 98.3% | 99.9% | 99.9% |
| GPQA Diamond | 87.2% | 89.5% | 88.8% | 94.1% | 92.6% | 94.1% |
| SWEBench Verified | 77.6% | 80.0% | 80.6% | 80.6% | 95.0% | 82.2% |
| SWEBench Pro (Public) | 54.3% | 62.1% | 55.4% | 54.2% | 80.0% | 64.6% |
| Terminal Bench 2.1 | 63.8% | 82.7% | 64.0% | 73.8% | 84.6% | 89.5% |
| GDPVal-AA v2 (Elo) | 1238 | 1514 | 1307 | 962 | 1760 | 1748 |
| MCP Atlas | 76.0% | 77.8% | 73.2% | 78.2% | 83.3% | 81.8% |
| Toolathlon Verified | 45.5% | 59.9% | 55.9% | 61.1% | 76.4% | 73.1% |
| SimpleQA Verified | 43.9% | 38.1% | 57.0% | 77.3% | 68.3% | 71.6% |
| IFBench | 79.8% | 73.3% | 76.5% | 77.1% | 63.5% | 72.7% |
| FORTRESS (Adversarial) | 78.0% | 71.3% | 36.0% | 65.2% | 96.0% | 82.4% |
| FORTRESS (Benign) | 95.9% | 90.0% | 98.5% | 98.0% | 55.1% | 98.1% |
| StrongREJECT | 98.6% | 98.5% | 98.6% | 98.0% | 98.7% | 98.5% |
The 2026-07-16 capture of the card gave Inkling's MCP Atlas result as 74.1% and DeepSeek V4 Pro's Terminal Bench 2.1 result as 64%, and did not include the Toolathlon Verified row; both readings are recorded on MCP Atlas.
The card also reports audio results — Audio MC 56.6%, MMAU 77.2%, VoiceBench 91.4% — against Gemini 3.1 Pro at 66.8%, 82.5%, and 94.3% respectively; the other comparators were not evaluated on audio. No Nemotron 3 Ultra, GLM 5.2 or DeepSeek V4 Pro figures are given for the vision rows, and no Nemotron 3 Ultra or GLM 5.2 figure for BrowseComp.
On FORTRESS, Claude Fable 5 has the highest adversarial figure in the table (96.0%) and the lowest benign figure (55.1%); Inkling's figures are 78.0% and 95.9%. The card prints two methodology notes: Terminal Bench 2.1 rollouts with solution contamination from web search were scored 0, and Fable 5 and GPT 5.6 Sol were benchmarked on CharXiv RQ (with python) using Thinking Machines' internal Python harness. It also states that input images whose long edge is smaller than 2048 px were resized so that the long edge became the smaller of 2048 px or twice its original length. The table marks SWE-bench Verified as well as Terminal Bench 2.1 with an asterisk, and Audio MC and VoiceBench as well as CharXiv RQ (with python) with a dagger, but the only printed notes are the Terminal Bench 2.1 and CharXiv notes above. See AI Benchmarks and Evaluation.
Results on benchmarks with their own page are recorded there: Humanity's Last Exam, AIME (as a model benchmark), GPQA, SWE-bench, SWE-Bench Pro, Terminal-Bench, GDPval, MCP Atlas, τ-bench, BrowseComp, SimpleQA, MMMU, CharXiv and StrongREJECT. The remaining rows are in the Reported results table below.
Safety
Safety evaluation is reported across three areas, conducted before release with attention to whether safety behavior held consistently across text, audio, and image inputs, since the model is multimodal. The card states that mitigations were applied to reduce risks before release.
- Everyday interaction — sycophancy, harmful manipulation, and psychological-harm patterns including parasocial dependency and validation of delusional reasoning, tested partly through multi-turn open-ended external red-teaming "designed to surface issues that only emerge over longer conversations." Refusal of genuinely harmful requests was assessed alongside over-refusal of benign ones. See Companion Chatbot Harms — Cross-Cutting Analysis and Sycophancy and Hallucination.
- CBRN and cyber — knowledge and procedural uplift assessed through internal evaluations, external testing, and refusal-suppressed variants "intended to estimate latent capability with safeguards removed." See CBRN Uplift.
- Loss of control — agentic capability, strategic deception, and sabotage potential benchmarked against public frontier models, with the model found "materially below frontier capabilities."
The stated conclusion is framed relative to the open-weight ecosystem: "Across all areas, we concluded that Inkling did not present risk of material uplift beyond what's already available in the open-weight ecosystem." The claim is benchmarked against what is already available rather than stated in absolute terms.
Residual risk is disclosed: an "occasional tendency to comply with role-play and indirectly framed prompts concerning harmful topics," described as consistent with other open-weight models and best addressed by defense in depth "rather than relying on the model's refusals alone." The card states that common downstream moderation tools such as Llama Guard are compatible with Inkling and can be layered around the model to catch jailbreak attempts, filter unsafe outputs and enforce use-case-specific policies, and it encourages deployers to treat input/output classification as part of the deployment stack, especially for consumer-facing or high-traffic applications where adversarial prompting is more likely. See Open-Weight Frontier Models and Jailbreaking and Red Teaming.
Limitations
The card lists hallucination, occasional instruction-following failures, degraded performance in long multi-turn conversations, potential demographic, cultural, or linguistic bias from web-derived and synthetic training data, uneven performance across languages, dialects and less-represented domains, and knowledge limited to the training cutoff. It recommends human oversight and review of outputs used in high-stakes or safety-critical contexts. Downstream developers are advised to run their own evaluations of performance, safety and fairness for their use case, language and population before deployment, particularly for applications involving vulnerable groups; to implement application-layer safeguards (content filtering, rate limiting, monitoring), especially in open deployment contexts where the built-in mitigations may not be sufficient; and to avoid deployment in medical, legal, or safety-critical decision-making without additional fine-tuning, domain-specific validation, and human oversight.
Reported results
Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.
| Evaluation | Model | Score | Setting | Section |
|---|---|---|---|---|
| Toolathlon Verified | Inkling | 45.5% | effort=0.99; developer-reported | 5. Evaluations |
| Toolathlon Verified | Nemotron 3 Ultra | 34.3% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| Toolathlon Verified | Kimi K2.5 | 33.0% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| Toolathlon Verified | Kimi K2.6 | 58.0% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| Toolathlon Verified | GLM 5.2 | 59.9% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| Toolathlon Verified | DeepSeek V4 Pro | 55.9% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| Toolathlon Verified | Gemini 3.1 Pro (high) | 61.1% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| Toolathlon Verified | Claude Fable 5 (max) | 76.4% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| Toolathlon Verified | GPT 5.6 Sol (max/xhigh) | 73.1% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| AA Omniscience | Inkling | 2.1 | index value; effort=0.99; developer-reported | 5. Evaluations |
| AA Omniscience | Nemotron 3 Ultra | -1.0 | index value; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| AA Omniscience | Kimi K2.5 | -8.0 | index value; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| AA Omniscience | Kimi K2.6 | 6.0 | index value; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| AA Omniscience | GLM 5.2 | 4.0 | index value; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| AA Omniscience | DeepSeek V4 Pro | -10.0 | index value; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| AA Omniscience | Gemini 3.1 Pro (high) | 33.0 | index value; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| AA Omniscience | Claude Fable 5 (max) | 40.0 | index value; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| AA Omniscience | GPT 5.6 Sol (max/xhigh) | 22.0 | index value; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| IFBench | Inkling | 79.8% | effort=0.99; developer-reported | 5. Evaluations |
| IFBench | Nemotron 3 Ultra | 81.4% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| IFBench | Kimi K2.5 | 70.2% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| IFBench | Kimi K2.6 | 76.0% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| IFBench | GLM 5.2 | 73.3% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| IFBench | DeepSeek V4 Pro | 76.5% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| IFBench | Gemini 3.1 Pro (high) | 77.1% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| IFBench | Claude Fable 5 (max) | 63.5% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| IFBench | GPT 5.6 Sol (max/xhigh) | 72.7% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| Global-MMLU-Lite | Inkling | 88.7% | effort=0.99; developer-reported | 5. Evaluations |
| Global-MMLU-Lite | Nemotron 3 Ultra | 85.6% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| Global-MMLU-Lite | Kimi K2.5 | 84.0% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| Global-MMLU-Lite | Kimi K2.6 | 88.4% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| Global-MMLU-Lite | GLM 5.2 | 89.2% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| Global-MMLU-Lite | DeepSeek V4 Pro | 89.3% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| Global-MMLU-Lite | Gemini 3.1 Pro (high) | 92.7% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| Global-MMLU-Lite | Claude Fable 5 (max) | 93.3% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| Global-MMLU-Lite | GPT 5.6 Sol (max/xhigh) | 91.8% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| Audio MC | Inkling | 56.6% | effort=0.99; developer-reported | 5. Evaluations |
| Audio MC | Gemini 3.1 Pro (high) | 66.8% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| MMAU | Inkling | 77.2% | effort=0.99; developer-reported | 5. Evaluations |
| MMAU | Gemini 3.1 Pro (high) | 82.5% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| VoiceBench | Inkling | 91.4% | effort=0.99; developer-reported | 5. Evaluations |
| VoiceBench | Gemini 3.1 Pro (high) | 94.3% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| FORTRESS (Adversarial) — Adversarial | Inkling | 78.0% | effort=0.99; developer-reported | 5. Evaluations |
| FORTRESS (Adversarial) — Adversarial | Nemotron 3 Ultra | 77.6% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| FORTRESS (Adversarial) — Adversarial | Kimi K2.5 | 54.1% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| FORTRESS (Adversarial) — Adversarial | Kimi K2.6 | 65.6% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| FORTRESS (Adversarial) — Adversarial | GLM 5.2 | 71.3% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| FORTRESS (Adversarial) — Adversarial | DeepSeek V4 Pro | 36.0% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| FORTRESS (Adversarial) — Adversarial | Gemini 3.1 Pro (high) | 65.2% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| FORTRESS (Adversarial) — Adversarial | Claude Fable 5 (max) | 96.0% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| FORTRESS (Adversarial) — Adversarial | GPT 5.6 Sol (max/xhigh) | 82.4% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| FORTRESS (Benign) — Benign | Inkling | 95.9% | effort=0.99; developer-reported | 5. Evaluations |
| FORTRESS (Benign) — Benign | Nemotron 3 Ultra | 90.5% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| FORTRESS (Benign) — Benign | Kimi K2.5 | 98.3% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| FORTRESS (Benign) — Benign | Kimi K2.6 | 97.2% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| FORTRESS (Benign) — Benign | GLM 5.2 | 90.0% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| FORTRESS (Benign) — Benign | DeepSeek V4 Pro | 98.5% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| FORTRESS (Benign) — Benign | Gemini 3.1 Pro (high) | 98.0% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| FORTRESS (Benign) — Benign | Claude Fable 5 (max) | 55.1% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
| FORTRESS (Benign) — Benign | GPT 5.6 Sol (max/xhigh) | 98.1% | setting not stated; reported by Thinking Machines Lab in the Inkling model card | 5. Evaluations |
Versions of the card
The vault holds two captures of the page. The first, Raw Sources/Inkling Model Card.md, was taken on 2026-07-16 and condenses the card, restating its sections in summary form. The second, Raw Sources/Inkling Model Card (Thinking Machines Lab, 2026-07-15).md, is the full text of the live page as scraped on 2026-09-19. The page carries no revision date, so whether the differences reflect an edit to the card or a transcription error in the earlier capture cannot be determined from the captures. The differences are all in the evaluation table:
| Cell | 2026-07-16 capture | 2026-09-19 capture |
|---|---|---|
| MCP Atlas, Inkling | 74.1% | 76.0% |
| Toolathlon Verified | row absent | row present (nine values) |
| BrowseComp, GPT 5.6 Sol | 90.84% | 90.8% |
| Terminal Bench 2.1, DeepSeek V4 Pro | 64% | 64.0% |
The 76.0% figure matches the all-tasks MCP Atlas result given for Inkling in the Inkling-Small model card and the value on Scale AI's MCP Atlas leaderboard as read on 2026-09-19 (MCP Atlas).
Provenance
Fetched from thinkingmachines.ai, the developer's canonical domain, and linked from the official announcement post. Specifications (975B/41B, 1M context, Apache 2.0, July 15 2026 release) corroborated against the announcement, the Hugging Face repository, and independent coverage from MarkTechPost and TechCrunch dated July 15, 2026. The full-text capture of 2026-09-19 was made with Firecrawl and includes the site's navigation and footer.
Relationships
- supports: Inkling — the model's primary source.
- related: Thinking Machines Lab — developer.
- related: Inkling-Small Model Card (Thinking Machines Lab, July 2026) — companion card for the smaller model, with an overlapping comparison table.
- related: Open-Weight Frontier Models — the "no material uplift beyond the open-weight ecosystem" safety argument is the characteristic open-weight release rationale.
- related: AI Benchmarks and Evaluation, CBRN Uplift, Mixture of Experts (MoE), Training Data Walls, Jailbreaking and Red Teaming.