Cohere has launched Parse 5, a document vision parsing model that turns unstructured images and documents into structured data for downstream AI agents and applications. The model supports OCR, tables, diagrams, images and visual grounding with bounding boxes in nine languages and can be deployed vi
Particle introduced Radar, a podcast search engine that transcribes and indexes spoken content for discovery by AI agents. The company says Radar has indexed more than 130,000 podcasts and is being used via API by hedge funds, AI search platforms and resellers.
A paper from Multiverse Computing introduces Quantization-Aware Healing (QAH), a distillation recipe that trains a 4-bit MXFP4 student against the original full-size teacher. Applied to GPT-OSS 120B→60B, the QAH model beats its bfloat16 recovered checkpoint on 7 of 9 benchmarks.
Granite 4.2 is a family of dense, decoder-only reasoning LLMs in three sizes (3B, 8B, 30B), pre-trained from scratch on ~15T tokens with a five-phase strategy and a 512K context extension. The 8B and 30B models undergo agentic RL for tool use; all models are Apache 2.0.
OpenAI is extending Codex-style agents through ChatGPT Work, a $20-per-month tier aimed at connecting LLMs to email, calendars and other SaaS tools. Engineers describe adoption gaps, permission hurdles and a debate over whether the model or the harness matters most.
Berd, launched in 2026, is a desktop app aimed at engineers and designers who want a persistent AI agent inside a local workspace. The Open Source Mac app links projects, skills, extensions, automations, model providers and local context.
Researchers at HumeAI report that some open-source ASR models reproduce benchmark-specific transcripts even when the audio contradicts them. Tests on VoxPopuli and LibriSpeech reveal dataset-linked acoustic cues, masked-number recovery, and orthographic switching.
Ramp has launched Router, an API-based model routing service that lets users switch among LLMs from OpenAI, Anthropic, xAI and others. The service is US-only, free through the end of 2026 with a $26 credit, and continues to charge for model inference.
Liquid AI released DSpark draft checkpoints for LFM2.5-1.2B-Instruct, LFM2.5-2.6B and LFM2.5-8B-A1B, reporting up to 3.18× GPU and 2.87× on-device throughput. The approach uses a small draft model and speculative decoding; checkpoints are available in Safetensors and GGUF.
LiquidAI published QAD Q4_0 GGUF checkpoints for LFM2.5-230M, -350M, -1.2B-Instruct and -2.6B, claiming recovery of most accuracy lost to 4-bit quantization. Benchmarks and throughput on real devices are reported alongside usage instructions for llama.cpp.