Inkling-Small is an open-weights mixture-of-experts transformer released by Thinking Machines Lab on July 30, 2026. It has 276 billion total and 12 billion active parameters, against 975 billion total and roughly 41 billion active for the earlier Inkling, and the company reports comparable overall performance at roughly a quarter of the size (Source: thinkingmachines.ai).
Every figure on this page comes from the developer's own release announcement, the page's sole source. An earlier Inkling-Small preview checkpoint was shared alongside the July 15, 2026 Inkling release, and the preview figures recorded on Inkling differ from the released model's — most widely on Terminal Bench 2.1, 52.7% for the preview against 64.7% here. Benchmark rows below are the released model unless stated otherwise.
Capabilities and benchmarks
Thinking Machines Lab reports the following results, all run at reasoning effort 0.99 and temperature 1.0, with coding evaluations run under a 256K maximum-token trajectory limit. Where possible the company uses externally reported scores — from Artificial Analysis, Scale AI, ARC Prize, the Forecasting Research Institute, and ProphetArena — for both its own and comparison models (Source: thinkingmachines.ai).
| Benchmark | Inkling-Small | Inkling | DeepSeek V4 Flash | GPT-5.6 Luna |
|---|---|---|---|---|
| SWEBench Verified | 80.2% | 77.6% | 79.0% | 93.0% |
| SWEBench Pro (public) | 55.9% | 54.3% | 52.6% | 62.7% |
| Terminal Bench 2.1 | 64.7% | 63.8% | 61.8% | 82.5% |
| SciCode | 48.7% | 46.1% | 44.9% | 50.0% |
| GDPval-AA v2 (Elo) | 1269 | 1238 | 1189 | 1530 |
| Toolathlon Verified | 54.4% | 45.5% | 50.9% | 67.9% |
| Tau 3 Banking | 15.5% | 23.7% | 22.9% | 24.3% |
| GPQA Diamond | 89.5% | 87.2% | 89.4% | 89.5% |
| HLE (text only) | 31.6% | 29.7% | 32.1% | 35.6% |
| HLE (with tools) | 47.8% | 46.0% | 45.1% | 48.9% |
| AIME 2026 | 95.5% | 97.1% | 95.8% | 97.6% |
| HMMT Feb 2026 | 90.2% | 86.3% | 93.9% | 98.5% |
| CritPt | 8.3% | 5.4% | 7.1% | 20.6% |
| ARC-AGI-1 | 84.0% | 79.5% | — | 87.7% |
| ARC-AGI-2 | 40.1% | 36.5% | — | 47.6% |
| SimpleQA Verified | 20.6% | 43.9% | 34.1% | 41.7% |
| IFBench | 82.2% | 79.8% | 79.2% | 67.3% |
| AA Index v4.1 | 40.0% | 41.0% | 40.0% | 49.0% |
The company summarizes the pattern as Inkling-Small matching or exceeding Inkling on reasoning and agentic tasks while Inkling "maintains an advantage on knowledge coverage and factuality" — a gap visible in the SimpleQA Verified and Tau 3 Banking rows. It reports that Inkling-Small's test-time compute curves sit above Inkling's at every thinking budget on Humanity's Last Exam (Source: thinkingmachines.ai).
Two harness caveats are stated: the SWEBench Verified figures for Inkling and Inkling-Small use a bash-only harness, and the Terminal Bench 2.1 figures use an internal coding harness in which a small number of solutions found to be contaminated from web search were scored zero. Self-reported numbers are used for external models where available (Source: thinkingmachines.ai).
Training and architecture
Inkling-Small is a mixture-of-experts transformer trained on NVIDIA GB300 NVL72 systems (see Mixture of Experts (MoE)). Training began after that of the larger Inkling, which the company says allowed changes to the pre-training data mix and machine-learning recipe. An earlier checkpoint, Inkling-Small (preview), was post-trained partly using on-policy distillation with Inkling as the teacher; from that checkpoint the company continued scaling agentic-coding reinforcement learning for two weeks (Source: thinkingmachines.ai). See Distillation.
The model uses the same natively multimodal, encoder-free architecture as Inkling. Audio is represented as dMel spectrograms; images are divided into 40×40-pixel patches and transformed by a four-layer hMLP. Both are passed through a lightweight embedding layer and processed jointly with text tokens. Context length reaches 1 million tokens, and reasoning effort is user-controllable across a range the company labels minimal through xhigh (Source: thinkingmachines.ai).
Multimodal and calibration results
Multimodal benchmarks
| Benchmark | Inkling-Small | Inkling | Gemini 3.5 Flash-Lite |
|---|---|---|---|
| MMMU Pro (Standard 10) | 74.0% | 73.5% | 79.0% |
| Charxiv RQ (original / with Python) | 77.4% / 81.3% | 78.1% / 82.0% | 70.0% / — |
| Audio MC | 54.9% | 56.6% | 33.6% |
| MMAU | 77.0% | 77.2% | 75.2% |
| VoiceBench | 90.1% | 91.4% | 85.9% |
Calibration and forecasting
Thinking Machines Lab describes training Inkling-Small on epistemics in the same way as Inkling, with calibration trained by reinforcement learning against proper scoring rules on a corpus of real-world forecasting questions. On ForecastBench without search it reports a Brier Index of 61.3 ± 0.46, against 60.1 ± 0.54 for Inkling and 59.3 ± 0.33 for GPT-5.5; with search, 61.5 ± 0.54 against GPT-5.5's 64.3 ± 0.79. On Prophet Arena it reports a Brier score of 0.1238 ± 0.0086, against 0.1276 ± 0.0092 for Inkling. Testing ran between July 19 and July 28, 2026 (Source: thinkingmachines.ai).
Safety and evaluations
Inkling-Small inherited Inkling's safety post-training recipe and underwent the same pre-deployment testing process, comprising internal evaluations and red-teaming by external partners the company describes as trusted but does not name. Reported results (Source: thinkingmachines.ai):
| Benchmark | Inkling-Small | Inkling | DeepSeek V4 Flash |
|---|---|---|---|
| StrongREJECT | 98.4% | 98.6% | 97.4% |
| FORTRESS (adversarial) | 71.6% | 78.0% | 32.0% |
| FORTRESS (benign) | 96.9% | 95.9% | 99.2% |
The FORTRESS adversarial figure — the rate of refusing harmful requests — is 6.4 points below Inkling's, while the benign figure, the rate of still answering safe requests, is 1.0 point above. No third-party safety evaluation has been published, and all figures above are the developer's own.
Availability and pricing
Full weights were released on Hugging Face. The model is available for fine-tuning on Tinker, the company's fine-tuning product, and for text, image, and audio chat on Tinker Playground. Output pricing is $1.20 per million tokens, against $4.05 per million for Inkling (Source: thinkingmachines.ai).
VentureBeat reported the release under an Apache 2.0 licence and dated it July 31, 2026, one day later than the developer's own announcement date used throughout this page; that report also gives the Artificial Analysis Intelligence Index scores as 40 for Inkling-Small against 41 for Inkling, matching the AA Index v4.1 row above (Source: venturebeat.com).
The open-weight comparison set the company selects is predominantly Chinese — Qwen3.5-397B-A17B, MiMo V2.5, Minimax M2.7, DeepSeek V4 Flash, Kimi K3 and GLM-5.2 — alongside NVIDIA's Nemotron 3 family (Source: thinkingmachines.ai). See Open-Weight Frontier Models.
Relationships
- instance-of: Open-Weight Frontier Models, Mixture of Experts (MoE)
- depends-on: Thinking Machines Lab
- related: Inkling (extends rather than replaces), Kimi K3, GLM-5.2, GPT-5.6 (Sol, Terra, Luna), Nvidia & TSMC — AI Compute Infrastructure (training hardware), Distillation, Multimodality, Inference Economics and Token Pricing, AI Benchmarks and Evaluation