AI developer tools · What shipped

BenchLM tracked 235 AI model releases in the 12 months ending October 8, 2026, averaging roughly one notable release every two days. OpenAI led with 21 tracked releases, followed by Alibaba (17), Google (16), and Anthropic (13). September 2026 was the busiest month with 42 releases tracked. Among classified releases, 46% were open-weight models, reflecting continued growth in openly available AI model releases across the industry.

BenchLM tracks 625 LLM benchmarks across 11 categories for 889 AI models as of October 2026, with frontier LLM token prices down 84% since March 2023. The platform recorded 235 AI model releases in the 12 months ending October 8, 2026—roughly one notable release every 2 days—and documents the largest tracked context window at 10M tokens from Pokee AI's Pokee-Isaac 28B. Of the top 50 AI models, 22% are open-weight, with Anthropic's Claude Opus 5.5 ranking as the #1 model overall with a score of 86.43/100.

oMLX 0.7.1.dev1 adds decision models, batched prefill for concurrent requests, and performance improvements for Qwen and GLM-5.3 models. Mean time to first token improvements include Qwen3.8-Flash-Next on M5 Max dropping from 2.08s to 1.81s (-13%), and GLM-5.3-Flash on M5 Max from 1.92s to 1.68s (-12%) with batched prefill. Decode throughput for Qwen3.8-Flash-Next increased from 39 to 91 tok/s (+133%) on M3 Ultra with group size 32 checkpoints. oMLX 0.7.0 brings faster Qwen3.8, GLM-5.3-Flash and MiMo V2 inference through prefill and decode optimizations including fused MoE kernels and exact QSA attention. Prompt processing improved by 33.5% at 32K context on Qwen3.8-Flash-Next. A completely rebuilt memory guard now uses configurable tiers (safe/balanced/aggressive) to manage RAM allocation, helping oMLX use more memory safely. oMLX 0.7.0rc1 demonstrates performance gains including Qwen3.8-Flash-Next prefill improvements of 31.9% at 16K tokens and 4-request batch decode throughput increases of 131.1% on Qwen3.8-27B with DFlash2. New model support includes Ternary Bonsai 2 with vision feature caching and MiMo V2.6 multimodal support with Lightning MTP and DFlash.