OpenRouter listed Ling 3.0 Flash VL (served by novita_labs): a 124B MoE with ~5.5B active params, native image/video input, visual agent capabilities, hybrid instant/reasoning modes, and tool calling, including a free route.
DeepSeek introduced DeepSeek-V4.1-Flash—the smallest model in its new family with native vision. It is a 552B MoE with a Causal Encoder–Decoder (8B active on input, 16B on output), claims benchmarks ahead of flagships including V4-Pro, shrinks KV cache to ~1/4 HBM and ~1/8 SSD vs prior gen, and published HF weights plus a tech report.
OpenAI Developers said GPT-Live-1 is available in the API: bring ChatGPT’s natural back-and-forth into your app with voice agents that listen while they speak, using the models and harness you choose. OpenAI’s Tibo also noted builders can use the same full-duplex voice stack shipped to ~1B ChatGPT users, with solid tool calling.
HuggingPapers reports NVIDIA published NVFP4-quantized Qwen3.8-Flash-Next weights on Hugging Face (125B MoE with hybrid attention), claiming ~63% smaller size with minimal accuracy loss. The official repo nvidia/Qwen3.8-Flash-Next-NVFP4 targets higher-throughput local/edge inference on NVIDIA hardware with a lower VRAM floor.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.