AI Model Intelligence Report — April 2026 (55+ LLMs Benchmarked)

[desiorac]

The Report

We just published the AI Model Intelligence Report for April 2026 — a comprehensive benchmark of 55+ open-weight LLMs across HumanEval, LiveCodeBench, SWE-bench, MMLU, and GSM8K.

Key Findings

• Kimi K2.5 leads SWE-bench at 76.8% — beating GPT-4o by 8.4 points
• Chinese labs hold 14 of top 20 spots — the landscape has shifted dramatically
• DeepSeek-R2 (74.2%) and Qwen3-72B (71.5%) round out the top 3
• Sub-7B models now reach 84%+ on MMLU — local deployment is increasingly viable
• Inference costs dropped 30x year-over-year for top-tier performance

What's in the Report

✅ Complete ranked table: all 55+ models, 5 benchmarks
✅ Per-model: size, license, context window, cost/M tokens
✅ Decision framework: coding assistants vs RAG vs agents vs local
✅ Trend analysis vs January 2026
✅ Chinese lab vs US lab capability gap breakdown

€9 one-time — no subscription.

👉 Landing page + checkout

Free preview: https://gist.github.com/desiorac/bbf02d04b58e361486b993df8bc06d1b