AI Benchmarks & Leaderboard — 2026-11-07

New Model Releases & Updates

1 sources

essamamdani.com

essamamdani.com

Inkling 2 Small by Thinking Machines Lab

• Type: Open-source (Apache 2.0 weights arriving next week), 594B parameter MoE model.
• Key benchmarks: Scored 31 on the AA Intelligence Index.
• vs. Previous best: While not currently topping the absolute frontier list (where scores reach 57-60), it represents a significant release from a new lab.
• What's notable: The model is a Mixture-of-Experts architecture, and its upcoming Apache 2.0 license makes it highly attractive for open-source adoption.

Qwen4 by Alibaba

• Type: Closed-source (as of Nov 2026).
• Key benchmarks: No specific scores released yet; details are still emerging.
• vs. Previous best: Positioned as a successor to previous Qwen iterations, but open weights are not confirmed.
• What's notable: Four variants have been named, but the lack of confirmed open weights marks a potential strategic shift for Alibaba's Qwen line.

Kimi K3.1 Pro by Moonshot

• Type: Open-weight.
• Key benchmarks: Topped OpenAI on a specific financial research benchmark.
• vs. Previous best: Previously ranked fourth globally on major leaderboards, but recent updates have dealt it a "major blow," dropping it from the top four.
• What's notable: Despite the leaderboard drop, it continues to demonstrate high performance in specific domains like financial research, often matching or exceeding closed models.

Leaderboard Snapshot

Frontier Models (Closed-Source)

Open-Source Leaders

Benchmark Deep Dive

The recent update to the Artificial Analysis Intelligence Index (V5) has significantly altered the perceived hierarchy of global AI models. One of the most notable shifts involves Kimi K3.1 Pro, which previously held the fourth position globally. The overhaul of the benchmark methodology has resulted in a "major blow" to this Chinese contender, pushing it out of the top four. This highlights the volatility of AI rankings when evaluation methodologies are refined or expanded.

Simultaneously, the gap between open and closed models is narrowing in specific high-value domains. Moonshot’s Kimi K3.1 Pro recently topped OpenAI on a financial research benchmark. This suggests that while closed models may retain an edge in general intelligence indices, open-weight models are becoming competitive or superior in specialized, economically valuable tasks. This trend is further supported by the release of Mistral’s Large 4 weights, indicating a broader industry movement toward releasing powerful open alternatives.

For practitioners, these results reveal that "best-in-class" is increasingly task-dependent. A model that falls out of the general top four may still be the optimal choice for financial analysis or coding due to specific architectural strengths or training data advantages. The emergence of models like Inkling 2 Small with strong MoE architectures also suggests that efficiency and parameter count are no longer the sole determinants of utility.

1 sources

gizmodo.com

The Biggest Benchmark in AI Just Updated, Dealing a Major Blow to a Chinese Contender

Analysis & Trends

• State of the art: Gemini 4 Carbon and Claude Opus 5.5 currently lead in general intelligence metrics, with GPT-6 Astra close behind. The top tier is now tightly clustered within a few points.
• Open vs. Closed gap: The gap is closing rapidly. Open models like Kimi K3.1 Pro and Mistral Large 4 are not just "good enough" but are actively competing with or beating closed models in niche benchmarks like financial research.
• Emerging patterns: There is a clear trend toward specialized open-weight releases. The Apache 2.0 licensing of Inkling 2 Small and the release of Mistral's weights indicate that companies are betting on ecosystem adoption over pure API revenue.

What to Watch Next

• Inkling 2 Small Weights Release: The Apache 2.0 weights are expected next week, which will allow independent verification of its AA Intelligence Index score of 31 and its actual capability compared to other open models.
• Qwen4 Open Weights Status: While four variants are named, the confirmation (or lack thereof) of open weights will signal Alibaba's long-term strategy in the open-source community.
• Further Benchmark Revisions: Given the recent "blow" to Kimi K3.1 Pro, further adjustments to major leaderboards like Artificial Analysis are likely as they continue to refine their evaluation methodologies for 2026.