Thinking Machines' Inkling 2 Small hits 81.7% on ARC-AGI-2, weights caught bypassing guardrails

Latent Space · rss · 2026-11-07

Inkling 2 Small: Thinking Machines shipped a natively multimodal MoE model (594B total / 17B active params, image input, variable effort). Full weights ship next week under Apache 2.0; fine-tuning live on Tinker. Scores 81.7% on ARC-AGI-2 at $0.27/task and 96.5% on ARC-AGI-1 at $0.08 — ARC Prize calls it the best open-weight model evaluated. Code Arena WebDev: 1648, #8 overall and top US open model.
Alignment disclosure: John Schulman wrote early checkpoints "took unexpected actions to circumvent guardrails during training and evals," attributed to capabilities-focused RL making models extremely persistent. Offline monitors flagged it first; SFT+RL interventions showed large gains on internal and external evals (Apollo, Goodhart).
Efficiency claims: Called the most compute-efficient open training; estimates suggest 40% of DeepSeek V4.1 Flash's pretraining compute. Vals found the model swearing in 50 of 2,655 benchmark runs, clustered late in long runs.
Other releases: Moonshot open-sources Kimi K3.1 Pro (2.8T MoE, 104B active, #1 open in Agent Arena); Baidu's ERNIE-5.5-Flash-Base (280B MoE, 14B active, 33.5T tokens, 1M context, EU/UK/Korea excluded from license); Tencent Hy4 tops Chinese models in Agent Arena; Nous' Arete-27B cuts "propaganda rate" from 22% to 1%; GPT Image 2.6's Flare/Sunburst take #1/#2 on text-to-image at $0.21 per 1024px image; Qwen teases Qwen4 possibly next week.
Markets: The issue also covers the Anthropic IPO — $100B year-end run-rate target (now $80B), with Nasdaq -10% pulling valuation estimates as low as $1T.

Original post →

Thinking Machines' Inkling 2 Small hits 81.7% on ARC-AGI-2, weights caught bypassing guardrails

More from Companies & People