Open-weight coding models close the gap with proprietary flagships, voice agents go mainstream, and video generation gets native audio.
Zhipu released GLM-5.3 with open weights under MIT, scoring 74.8% on SWE-bench — within a point of closed frontier models — at roughly one-fifth the API cost.
Veo now generates synchronized dialogue, sound effects, and ambience directly with video output, removing the need for a separate audio pass.
Moonshot Kimi k3 extends to 256K tokens with strong retrieval fidelity, making single-pass analysis of full codebases and legal documents practical.
The new release adds local LoRA fine-tuning from a single CLI command, lowering the barrier for domain-specific local models on consumer GPUs.
ElevenLabs streaming voice API now sustains sub-300ms round-trip latency, enabling natural phone-grade conversational agents.
Fable targets generative UI and design iteration loops, with state-of-the-art scores on multi-step creative agent benchmarks.
Runway opened Act-One performance capture to enterprise plans with team workspaces, SSO, and per-seat billing.
Black Forest Labs shipped a finetune that dramatically improves multi-line text in generated images — a long-standing pain point for marketing assets.
New quantization research cuts 70B-class model requirements to 16GB VRAM with negligible quality loss, reviving high-end consumer local inference.
At $0.10/$0.40 per million tokens with 1M context and top-tier latency, Gemini 3.7 Lite undercuts every high-volume classification and RAG alternative.
New editions are published every Sunday morning. Subscribe to our free newsletter or feed to never miss a breakthrough.
