Tag: #vision18 results
Anthropic has launched Claude Opus 4.7, significantly enhancing software development capabilities to a 'hands-off reliable' level for challenging coding tasks compared to Opus 4.6. Image recognition resolution exceeds 3x the previous generation. Improvements also include better instruction fidelity and stability for long-duration tasks.
Microsoft launched the open-weight Phi-4 15B multimodal model, which uniquely decides when to deeply reason or directly answer. With 15 billion parameters, it's designed for image description, UI element detection, and complex math reasoning tasks.
Chinese AI startup DeepSeek has added multimodal capabilities to its flagship chatbot, allowing it to process images and videos alongside text. This first-time feature is in limited release to select users, shortly after launching V4 model with price cuts. It positions DeepSeek competitively with rivals offering vision functions.
Together AI has expanded its fine-tuning service with native support for tool calling, reasoning, and vision-language models. It now enables training of 100B+ parameter models, offers up to 6× higher throughput, and provides job cost and ETA estimates.
Developer releases abliterated Qwen3.5-9B with vision, achieving 0% refusal rate via two-stage orthogonal projection + LoRA method. Outperforms heretic version's 46% refusals. Available on Ollama and Hugging Face model card.
A Reddit post details how to enable multimodal image understanding in Qwen 3.5 using llama.cpp via opencode. Users add a 'modalities' configuration to the opencode.json file specifying text and image inputs. This allows local running of the 35B model with vision capabilities.
Meituan advances native multimodal by treating images and speech as tokens for prediction. Approach claims discrete vision has no performance ceiling. Bold strategy for unified token modeling.
AWS integrates Amazon S3 general purpose buckets with SageMaker Unified Studio for easy use of unstructured data in ML workflows. This post demonstrates fine-tuning Llama 3.2 11B Vision Instruct for visual question answering using SageMaker Catalog. It simplifies LLM training with S3-stored data.
PKU Peng Yuxin team’s TARA aligns multimodal representations with taxonomy knowledge for hierarchical visual recognition. Boosts HCA 3-6% on iNat21, strong on unknown species via TerraIncognita. Improves VQA and linear probing; CVPR 2026.
Experiment tested 4 frontier multimodal models on valuing 15 paintings worth $1.46B using image-only or image + metadata. Found 'recognition vs commitment gap': models recognize art visually but hesitate on valuation without text. Gemini 3.1 Pro led both settings; GPT-5.4 surged with metadata.