there's a new multimodal retrieval model in town 🤠 @llama_index released vdr-2b-multi-v1 > uses 70% less image tokens, yet outperforming other dse-qwen2 based models > 3x faster inference with less VRAM 💨 > shrinkable with matryoshka 🪆 > can do cross-lingual retrieval! https://t.co/d1DyJcEHbN
🤖 Ultravox is a multimodal LLM that processes text and speech directly, eliminating the need for separate ASR. Takes in audio and emits streaming text. 🤯 The time-to-first-token (TTFT) is approx. 150ms. Tokens-per-second rate of ~60 using a Llama 3.1 8B backbone. https://t.co/90bK82v4Q5
New model added to the leaderboard! Model Name https://t.co/kXRsOhPEIO Overall rank: 2810 Rank in 1.5B category: 360 Benchmarks Average: 4.53 IFEval: 16.06 BBH: 2.17 MATH Lvl 5: 0.0 GPQA: 1.45 MUSR: 5.71 MMLU-PRO: 1.8
LocalAI has announced several new models aimed at enhancing artificial intelligence capabilities. The latest addition is 'nightwing3-10b-v0.1', a model based on Falcon3-10B, which can be initiated with the command `local-ai run nightwing3-10b-v0.1`. Another introduction is 'negative_llama_70b', designed for strong roleplay and creative writing, accessible via `local-ai run negative_llama_70b`. Additionally, 'finemath-llama-3b' is a continual pre-trained model focused on math and web datasets, which can be run using `local-ai run finemath-llama-3b`. The company also released 'LocalAI-functioncall-phi-4-v0.1' and its updated version 'LocalAI-functioncall-phi-4-v0.2', both of which feature function call capabilities. Furthermore, a new multimodal retrieval model named 'vdr-2b-multi-v1' has been released by @llama_index, boasting 70% less image token usage and three times faster inference. This model enables cross-lingual retrieval and is designed to be shrinkable with matryoshka technology. In addition, several models have been added to the leaderboard, with varying ranks and benchmark scores, indicating their performance across different categories.