Key takeaway
Thinking Machines Lab released Inkling 2 Small on November 5, 2026: a mixture-of-experts model with 594B total parameters and 17B active parameters, image input and variable effort, scoring 31 on the Artificial Analysis Intelligence Index. It ships first as a closed-weight model on Tinker, with full weights planned for Apache 2.0 release the following week after external safety testing and red-teaming.
What happened
Thinking Machines Lab published Inkling 2 Small on its official site on November 5, 2026. The company describes it as its second model generation, positioned as stronger on long-horizon autonomous work and easier to use inside real agent loops.
The release follows three stated steps. First, the model is available through Tinker with closed weights, following the company's stated path to open weights. Second, external safety testing and red-teaming are completed. Third, the company intends to release the full weights under an Apache 2.0 license the following week, with checkpoints downloadable at the same time. The company also states that the weight release remains subject to further safety evaluations by it and its partners.
Key facts
| Item | Company statement |
|---|---|
| Architecture | Mixture-of-Experts (MoE) |
| Parameters | 594B total, 17B active |
| Inputs | Text and images |
| Reasoning effort | Variable effort supported |
| Vision encoder | 292M parameters based on Perception Encoder; 20x20-pixel patches with 2x2 spatial pooling |
| Intelligence index | Artificial Analysis Intelligence Index 31 |
| Initial access | Tinker, closed weights; available for trial, fine-tuning and sampling on Tinker Playground |
| Weight plan | Apache 2.0 release the following week, with checkpoints |
| Design ranking | DesignArena Website score 1,340, third place |
On DesignArena's November 5, 2026 Website leaderboard snapshot, Inkling 2 Small ranks third at 1,340, behind Muse Wonder (max) at 1,366 and Muse Spark 1.3 at 1,351, and ahead of GPT-6 Astra (max) at 1,333, Kimi K3 at 1,328 and Claude Opus 5.5 at 1,321. The company says that pending the weight release, the model is set to be the top open-weights model on DesignArena.
Benchmark results
The company published results across coding, vision, knowledge work and long-context reasoning, alongside comparison models.
| Benchmark | Inkling 2 Small | Company annotation |
|---|---|---|
| FrontierCode v2.0 | 25.4% | Second-highest in row; highest is DeepSeek-V4.1-Flash-1017 at 28.5% |
| DeepSWE v1.1 | 72.1% | Second-highest; highest is DeepSeek-V4.1-Flash-1017 at 77.0% |
| SWEBench Pro v2 (Hard) | 86.3% | Second-highest; highest is Kimi K3 at 88.2% |
| Vibe Code Bench v1.1 | 81.6% | Below DeepSeek-V4.1-Flash-1017 at 87.7% and Kimi K3 at 85.0% |
| PostTrainBench v1.2 | 31.9% | Second-highest; highest is GLM-5.3-Flash at 36.8% |
| Terminal Bench Science v0.1 | 5.7% | Below DeepSeek-V4.1-Flash-1017 at 12.9% and Kimi K3 at 8.1% |
| PMPP-Hard (69 tasks, 10 hours) | 54.25/69, AUC 494.27 | Below GPT-6 Astra at 63/69, Kimi K3.1 Pro at 60.5/69 and Claude Opus 5.5 at 60/69 |
| Code Arena (WebDev) | 1,653 | Second-highest; highest is DeepSeek-V4.1-Flash-1017 at 1,665 |
| AA-3D | 1,021 | Highest in row; next is Kimi K3 at 982 |
| AA-Briefcase v1.1 | 1,527 | Second-highest; highest is DeepSeek-V4.1-Flash-1017 at 1,570 |
| GDPVal-AA v2.1 | 1,630 | Below DeepSeek-V4.1-Flash-1017 at 1,746 |
| JobBench | 70.5% | Highest in row |
| Vals Finance Index | 61.5% | Highest in row; next is DeepSeek-V4.1-Flash-1017 at 60.9% |
| HLE-Diamond | 28.7% | Highest in row |
| AA-LCR v1.1 | 83.3% | Second-highest; highest is Kimi K3 at 88.7% |
| LongBench v2 | 70.6% | Highest in row |
| AA-Omniscience v1.1 | 16.1 / 73.3 non-hallucination rate | Score below Muse Spark Open at 25.8; non-hallucination rate above its 72.0 |
On individual vision tests, the company reports ScreenSpot Pro at 83.8% (91.5% with Python), CharXiv RQ at 85.6% (87.7% with Python), Chartography at 36.7% (80.4% with Python), MMMU Pro (Standard 10) at 81.6% and PerceptionBench at 64.7%, marked as highest or second-highest in their rows.
Autonomous research demo
The company documents an end-to-end run to show long-horizon autonomy: Inkling 2 Small autonomously trained Qwen3.5-4B on Tinker to forecast time series across electricity demand, wind and solar generation, weather, air quality and Wikipedia page views.
Per the published trace, the model wrote its own evaluation script, built 12,500 training windows, ran several supervised fine-tuning rounds, and tried one reinforcement-learning phase using the task score as reward. Forecasts became overconfident and the score fell from 26% to 23%, so the model stopped that run; it submitted the checkpoint from its fourth supervised fine-tuning run.
The run took 4 hours 17 minutes and spent $426 against a $500 budget. On unseen August-September data, the final model produced about 15% less error than the base model.
Safety and the open-weights plan
The company says this release had broader safety evaluation coverage than any previous one: alongside the release checkpoint it trained an internal helpful-only variant, ran both through a wider evaluation suite, and involved more external testers.
Published numbers include a FORTRESS average risk score (ARS, lower is better) of 16.2, an over-refusal score (ORS) of 20.4, DistressBench 82.0, StrongREJECT 97.1, PatchEval 79.6 and Vals CyberBench 72.7. The company states that on chemical, biological, radiological and nuclear (CBRN) risk, knowledge and capability scores sit below the open-weight frontier; on offensive and dual-use agentic cyber benchmarks it improves substantially over the previous generation but remains behind the open-weight frontier. It also notes the model sometimes overreacts with disclaimers in ordinary emotional-support queries.
On weights, the company states that release depends on further safety evaluations by it and its partners, after which the weights ship under Apache 2.0 with downloadable checkpoints.
Industry impact
Two signals stand out in the public material. The first is capability placement: across coding benchmarks Inkling 2 Small mostly sits in the second tier behind DeepSeek-V4.1-Flash-1017 and Kimi K3, while it takes the top score in several knowledge-work and finance rows, including JobBench at 70.5%, Vals Finance Index at 61.5% and LongBench v2 at 70.6%. The second is licensing: if the Apache 2.0 release lands next week as planned, the open-weights camp gains a model at the 594B-total-parameter scale, where few open models currently exist.
The distinction that matters: "next week" is a stated plan, not a completed fact. The company makes external safety testing and red-teaming a precondition and leaves room to adjust timing based on results. That is this site's reading of the announcement.
A further observable point is customizability. The company reports that on one public financial-document summarization task, Inkling 2 Small reached the previous generation's final performance in 75 training steps, which sets the usable range of per-run training cost for enterprises fine-tuning on private data.
Official source
- Thinking Machines Lab: Introducing Inkling 2 Small (November 5, 2026)