🚀 RELEASES & LAUNCHES
DeepSeek releases V4 Flash Vision Exp with image understanding
DeepSeek released V4 Flash Vision Exp, an experimental multimodal model.- Targets developers building coding and multimodal agent workflows.
- Adds JPEG, PNG, GIF, and WebP image input while retaining V4 Flash text capabilities.
- Supports a 1M-token context window and caps each image at 384 tokens after resizing.
- DeepSeek reports multimodal-agent performance near Claude Opus 4.8; third-party reporting says it led V4 Flash on six of seven text-agent benchmarks.
- Off-peak pricing is $0.22 per million input tokens and $0.66 per million output tokens, matching V4 Flash.
Pipecat releases PhoneLLM Alpha 1 for low-latency voice agents
Pipecat released PhoneLLM Alpha 1, an open-weights model for voice agents.- Targets voice agents for financial services, healthcare, retail, hospitality, and customer service.
- Fine-tuned NVIDIA Nemotron 3 Nano 30B-A3B with 30B total parameters and 3.5B active parameters.
- PhoneBench score reached 72.3%, with about 600ms P95 time-to-first-token and estimated cost of $0.0025 per minute.
- Released under the BSD 2-Clause license on Hugging Face, with vLLM and SGLang deployment support.
- The model is trained for low-latency conversational tool use without extended reasoning, enabling self-hosted voice-agent deployments.