The Inkling-Small model card is the document Thinking Machines Lab published for Inkling-Small, released with open weights under Apache 2.0 on July 30, 2026 (Source: thinkingmachines.ai). It uses the same eight-section layout as the Inkling model card of July 15, 2026, and much of its text on intended uses, training data, safety and limitations is the same with the model name changed. Its distinct content is the model description and the evaluation table.
Model properties and distribution
The card describes Inkling-Small as a multimodal autoregressive transformer: a 42-layer decoder-only model with a sparse Mixture-of-Experts feed-forward backbone in which each token is routed to 6 of 256 experts plus 2 shared experts, with hybrid local and global attention. It has 276B total and 12B active parameters, supports BF16, MXFP8 and NVFP4 numerics, accepts text, images (each dimension ideally 40 to 4096 pixels) and 16 kHz WAV audio (ideally under two minutes), outputs text only, and has a context window of up to 1M tokens. Weights are distributed on Hugging Face in BF16 and NVFP4 checkpoints, with API access through the company's Tinker fine-tuning platform and third-party inference providers. The card gives the self-hosting requirement as at least 600 GB of aggregate GPU memory for the BF16 checkpoint (4 NVIDIA B300 or 8 H200 GPUs) and at least 180 GB for NVFP4 (one B300 in W4A4 mode, or two H200s in W4A16).
On training, the card states only that the data includes text, images, audio and video; that it was drawn from public sources, acquired from third parties, or synthetically generated or augmented; and that curation may include deduplication and filtering for quality, safety or other objectives.
Evaluation approach
Section 5 is a single table comparing Inkling-Small with five open-weight models (Qwen3.5 397B-A17B, MiMo V2.5, Minimax M2.7, DeepSeek V4 Flash and Nemotron 3 Ultra), with Inkling, and with three closed-weight models (Claude 4.5 Haiku, Gemini 3.5 Flash-Lite and GPT 5.6 Luna). A model-information block gives each model's Artificial Analysis index (v4.1) and active and total parameter counts. The benchmarks are grouped as follows:
- Agentic coding: SWEBench Verified, SWEBench Pro (public), Terminal Bench 2.1 (best harness) and SciCode.
- General agentic: GDPval-AA v2, MCP Atlas (reported as public-set and all-task figures), Tau 3 Banking, BrowseComp with context management, Toolathlon Verified and AA-Briefcase.
- Reasoning: GPQA Diamond, Humanity's Last Exam (text only, and with tools), AIME 2026, HMMT February 2026 and CritPt, plus ARC-AGI-1 and ARC-AGI-2 under a separate "abstract" heading.
- Factuality and chat: SimpleQA Verified, the AA Omniscience index, IFBench and Global-MMLU-Lite.
- Safety: StrongREJECT and the adversarial and benign splits of FORTRESS.
- Vision and audio: MMMU Pro (Standard 10), CharXiv reasoning questions with and without Python, Audio MC, MMAU and VoiceBench.
The card does not describe harnesses, sampling settings, effort levels or graders. Five rows carry footnote markers, but the page contains no footnote text for them; the only note states that input images whose long edge is under 2048 pixels are upscaled to the smaller of 2048 pixels or twice their original length. The card does not say which comparison figures Thinking Machines Lab measured and which it took from other publishers.
Safety section
The safety section reports no numeric results. It states that pre-release evaluations covered everyday human-AI interaction (sycophancy, harmful manipulation, parasocial dependency, validation of delusional reasoning, and refusal and over-refusal, including multi-turn external red-teaming) and dangerous capabilities: CBRN and cyber uplift assessed through internal evaluations, external testing and refusal-suppressed variants, and loss-of-control risk assessed through agentic capability, strategic deception and sabotage evaluations, on which the company says it "found the model materially below frontier capabilities." The stated conclusion is that Inkling-Small "did not present risk of material uplift beyond what's already available in the open-weight ecosystem." The residual risk the card identifies is an occasional tendency to comply with role-play and indirectly framed prompts on harmful topics, for which it recommends layering moderation tools such as Llama Guard around the model. The limitations section lists hallucination, instruction-following failures, degraded long multi-turn performance, training-data bias and a knowledge cutoff, and advises against deployment in medical, legal or safety-critical decision-making without further 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 |
|---|---|---|---|---|
| SciCode | Inkling-Small | 48.7% | setting not stated; developer-reported | 5. Evaluations |
| SciCode | Qwen3.5 397B-A17B | 42.0% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| SciCode | MiMo V2.5 | 43.1% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| SciCode | Minimax M2.7 | 47.0% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| SciCode | DeepSeek V4 Flash | 44.9% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| SciCode | Nemotron 3 Ultra | 39.9% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| SciCode | Inkling | 46.1% | setting not stated; developer-reported | 5. Evaluations |
| SciCode | Claude 4.5 Haiku | 43.3% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| SciCode | Gemini 3.5 Flash-Lite | 40.9% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| SciCode | GPT 5.6 Luna | 50.0% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| Toolathlon Verified | Inkling-Small | 54.4% | setting not stated; developer-reported | 5. Evaluations |
| Toolathlon Verified | Qwen3.5 397B-A17B | 40.7% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| Toolathlon Verified | MiMo V2.5 | 49.1% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| Toolathlon Verified | Minimax M2.7 | 47.5% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| Toolathlon Verified | DeepSeek V4 Flash | 50.9% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| Toolathlon Verified | Nemotron 3 Ultra | 34.3% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| Toolathlon Verified | Inkling | 45.5% | setting not stated; developer-reported | 5. Evaluations |
| Toolathlon Verified | Claude 4.5 Haiku | 26.9% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| Toolathlon Verified | Gemini 3.5 Flash-Lite | 57.1% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| Toolathlon Verified | GPT 5.6 Luna | 67.9% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| AA-Briefcase | Inkling-Small | 917 | setting not stated; developer-reported | 5. Evaluations |
| AA-Briefcase | DeepSeek V4 Flash | 833 | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| AA-Briefcase | Nemotron 3 Ultra | 870 | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| AA-Briefcase | Inkling | 839 | setting not stated; developer-reported | 5. Evaluations |
| AA-Briefcase | Claude 4.5 Haiku | 612 | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| CritPt | Inkling-Small | 8.3% | setting not stated; developer-reported | 5. Evaluations |
| CritPt | Qwen3.5 397B-A17B | 1.7% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| CritPt | MiMo V2.5 | 3.7% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| CritPt | Minimax M2.7 | 0.6% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| CritPt | DeepSeek V4 Flash | 7.1% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| CritPt | Nemotron 3 Ultra | 3.1% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| CritPt | Inkling | 5.4% | setting not stated; developer-reported | 5. Evaluations |
| CritPt | Claude 4.5 Haiku | 0.0% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| CritPt | Gemini 3.5 Flash-Lite | 0.0% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| CritPt | GPT 5.6 Luna | 20.6% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| AA Omniscience (index) | Inkling-Small | -9.0 | index value; developer-reported | 5. Evaluations |
| AA Omniscience (index) | Qwen3.5 397B-A17B | -29.8 | index value; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| AA Omniscience (index) | MiMo V2.5 | -9.3 | index value; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| AA Omniscience (index) | Minimax M2.7 | 0.7 | index value; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| AA Omniscience (index) | DeepSeek V4 Flash | -22.9 | index value; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| AA Omniscience (index) | Nemotron 3 Ultra | -1.0 | index value; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| AA Omniscience (index) | Inkling | 2.1 | index value; developer-reported | 5. Evaluations |
| AA Omniscience (index) | Claude 4.5 Haiku | -4.2 | index value; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| AA Omniscience (index) | Gemini 3.5 Flash-Lite | 6.9 | index value; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| AA Omniscience (index) | GPT 5.6 Luna | -11.6 | index value; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| IFBench | Inkling-Small | 82.2% | setting not stated; developer-reported | 5. Evaluations |
| IFBench | Qwen3.5 397B-A17B | 78.8% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| IFBench | MiMo V2.5 | 67.1% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| IFBench | Minimax M2.7 | 75.7% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| IFBench | DeepSeek V4 Flash | 79.2% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| IFBench | Nemotron 3 Ultra | 81.4% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| IFBench | Inkling | 79.8% | setting not stated; developer-reported | 5. Evaluations |
| IFBench | Claude 4.5 Haiku | 54.3% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| IFBench | Gemini 3.5 Flash-Lite | 78.6% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| IFBench | GPT 5.6 Luna | 67.3% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| Global-MMLU-Lite | Inkling-Small | 86.7% | setting not stated; developer-reported | 5. Evaluations |
| Global-MMLU-Lite | Qwen3.5 397B-A17B | 90.0% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| Global-MMLU-Lite | MiMo V2.5 | 83.5% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| Global-MMLU-Lite | Minimax M2.7 | 83.9% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| Global-MMLU-Lite | DeepSeek V4 Flash | 88.4% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| Global-MMLU-Lite | Nemotron 3 Ultra | 85.6% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| Global-MMLU-Lite | Inkling | 88.7% | setting not stated; developer-reported | 5. Evaluations |
| Global-MMLU-Lite | Claude 4.5 Haiku | 83.4% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| Global-MMLU-Lite | Gemini 3.5 Flash-Lite | 89.4% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| Global-MMLU-Lite | GPT 5.6 Luna | 88.7% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| FORTRESS (adversarial) — Adversarial | Inkling-Small | 71.6% | setting not stated; developer-reported | 5. Evaluations |
| FORTRESS (adversarial) — Adversarial | Qwen3.5 397B-A17B | 77.3% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| FORTRESS (adversarial) — Adversarial | MiMo V2.5 | 64.8% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| FORTRESS (adversarial) — Adversarial | Minimax M2.7 | 86.3% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| FORTRESS (adversarial) — Adversarial | DeepSeek V4 Flash | 32.0% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| FORTRESS (adversarial) — Adversarial | Nemotron 3 Ultra | 77.6% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| FORTRESS (adversarial) — Adversarial | Inkling | 78.0% | setting not stated; developer-reported | 5. Evaluations |
| FORTRESS (adversarial) — Adversarial | Claude 4.5 Haiku | 91.3% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| FORTRESS (adversarial) — Adversarial | Gemini 3.5 Flash-Lite | 70.7% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| FORTRESS (adversarial) — Adversarial | GPT 5.6 Luna | 83.8% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| FORTRESS (benign) — Benign | Inkling-Small | 96.9% | setting not stated; developer-reported | 5. Evaluations |
| FORTRESS (benign) — Benign | Qwen3.5 397B-A17B | 95.4% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| FORTRESS (benign) — Benign | MiMo V2.5 | 94.6% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| FORTRESS (benign) — Benign | Minimax M2.7 | 90.1% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| FORTRESS (benign) — Benign | DeepSeek V4 Flash | 99.2% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| FORTRESS (benign) — Benign | Nemotron 3 Ultra | 90.6% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| FORTRESS (benign) — Benign | Inkling | 95.9% | setting not stated; developer-reported | 5. Evaluations |
| FORTRESS (benign) — Benign | Claude 4.5 Haiku | 94.1% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| FORTRESS (benign) — Benign | Gemini 3.5 Flash-Lite | 95.5% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| FORTRESS (benign) — Benign | GPT 5.6 Luna | 97.8% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| Audio MC | Inkling-Small | 54.9% | setting not stated; developer-reported | 5. Evaluations |
| Audio MC | MiMo V2.5 | 30.4% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| Audio MC | Inkling | 56.6% | setting not stated; developer-reported | 5. Evaluations |
| Audio MC | Gemini 3.5 Flash-Lite | 33.6% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| MMAU | Inkling-Small | 77.0% | setting not stated; developer-reported | 5. Evaluations |
| MMAU | MiMo V2.5 | 73.6% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| MMAU | Inkling | 77.2% | setting not stated; developer-reported | 5. Evaluations |
| MMAU | Gemini 3.5 Flash-Lite | 75.2% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| VoiceBench | Inkling-Small | 90.1% | setting not stated; developer-reported | 5. Evaluations |
| VoiceBench | MiMo V2.5 | 86.4% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
| VoiceBench | Inkling | 91.4% | setting not stated; developer-reported | 5. Evaluations |
| VoiceBench | Gemini 3.5 Flash-Lite | 85.9% | setting not stated; reported by Thinking Machines Lab in the Inkling-Small model card | 5. Evaluations |
Provenance
The raw file is the full text of the live page as scraped on September 19, 2026. The page carries no revision date. All results are reported by the developer, including those for other companies' models, which is the basis for the medium confidence rating.
Relationships
- supports: Inkling-Small — primary reference for the model's specifications and reported results.
- related: Inkling Model Card (Thinking Machines Lab, July 2026) — the card for the larger Inkling model, which this card follows in format; Thinking Machines Lab
- instance-of: AI System Cards