Proximal has released Inkling 2 Small, a Mixture-of-Experts model with 594B total parameters and 17B active, positioning it as a frontier-class model with vision, agentic coding, and variable-effort reasoning. The model scores 31 on the Artificial Analysis Intelligence Index, which the company says far surpasses its first-generation models. It launches on Proximal's Tinker platform with closed weights, with an Apache 2.0 open-weight release planned after external safety testing and red-teaming. The benchmark picture is honest but complex. On agentic coding tasks, Inkling 2 Small lands in a tight pack with DeepSeek-V4.1-Flash and Kimi K3, rarely claiming the top score outright. On SWEBench Pro v2 Hard, it scores 86.3% against Kimi K3's 88.2%. On Vibe Code Bench, it posts 81.6% while DeepSeek leads at 87.7%. The pattern repeats across most tables: competitive but not dominant. Where Inkling 2 Small does lead is in vision benchmarks — GDP.pdf (25.2%), MMMU Pro (81.6%), PerceptionBench (64.7%), and Humanity's Sixth Sense (33.3%) all show top marks. The kernel optimization results on PMPP-Hard tell a more revealing story. GPT-6 Astra leads with 616.3 AUC task-hours and 63 mean tasks solved. Inkling 2 Small sits fifth at 494.27 AUC with 54.25 mean solved, behind Kimi K3.1 Pro, Claude Opus 5.5, and DeepSeek V4.1 Flash. The evaluation methodology is commendable — four independent attempts, 10-hour budgets, dual LLM audits with third-review tiebreaking — but the positioning is clearly mid-pack on the hardest autonomous work. The autoresearch demonstration is the most interesting technical claim. Inkling 2 Small autonomously trained a LoRA on Qwen3.5-4B for time-series forecasting, spending $426 in Tinker credits over 4.5 hours and more than doubling the forecasting score. This is a model training another model on a commercial API with a budget constraint — a concrete demonstration of recursive AI labor, not just a benchmark number. The architecture choices are deliberate. A 292M-parameter vision encoder based on Perception Encoder uses 20×20-pixel patches with 2×2 spatial pooling, interleaving image patches with text tokens. Training included RL on hundreds of thousands of long-horizon coding tasks with agentic RL graders evaluating approach quality, not just unit-test pass rates. Early checkpoints were tested on OpenCode under the name Fledge-Alpha. The open-weight commitment is the real structural story. Proximal says it will release full weights under Apache 2.0 after external safety testing — following a staged approach they call a 'safe path to open weights.' If they follow through, this puts a 594B MoE model into the open ecosystem, which changes the competitive dynamics for every lab building on closed APIs. The model is already available for fine-tuning and sampling on Tinker. What Proximal has built is a solid second-generation model that competes in the frontier pack without clearly leading it. The generative value lies less in the model itself — which sits in a crowded field — and more in the open-weight release and the Tinker ecosystem play, which could democratize access to MoE-class capabilities if the Apache 2.0 promise materializes.
