Descriptions:
Fahd Mirza’s weekly AI recap covers a packed release cycle headlined by Claude Opus 4.8 from Anthropic — same price as 4.7 but featuring dynamic workflows in research preview that let the model spawn hundreds of parallel sub-agents in a single session. Fast Mode is 2.5x faster and 3x cheaper, and early testers report the model flags its own uncertainties four times more often than 4.7.
Other notable releases include MiniMax’s upcoming M3 model with sparse attention (9.7x faster pre-fill, 15.6x faster decoding versus M2), Step 3.7 Flash from Depth One (198B total parameters, 11B active, up to 400 tokens/second, second place on SWEBench Pro at 56.3), and MiniCPM’s 1B on-device model with a toggleable think mode. Liquid AI’s LFM hybrid (8.3B total, 1.5B active) supports nine languages and blends recurrent layers with attention for strong CPU and GPU throughput. Bonsai Image compresses a Flux 4B transformer to just 0.93GB using one-bit weights — down from 7.7GB — while Stable Audio 3 generates tracks up to 6 minutes 20 seconds across 80-plus languages.
The recap rounds out with DS4, a native DeepSeek V4 Flash inference engine with persistent disk KV caching, Microsoft Lense (3.8B text-to-image, tested poorly in hands-on use), a 35B scientific multimodal model from Shanghai AI Lab, and Long Cat, an open-source audio-driven avatar video generator.
📺 Source: Fahd Mirza · Published May 30, 2026
🏷️ Format: Roundup
