Topic
AI models combining a vision encoder with a language model to generate text descriptions, answers, or analysis from images.
Nemotron 3.5 Super VL, NVIDIA's 120B model, had GGUF, MLX, FP8 and EXL3 conversions on Hugging Face a day after its October 15 release. The roughly 128 GB memory figure comes from a DEV Community write-up whose author did not run the model.
Reality
Researchers posting on LessWrong fine-tuned 15 vision-language models on narrow image tasks and saw unsafe behavior spread to unrelated tasks. The failures, including risky phone-agent actions, all showed up on tasks outside the training data, where a test suite built for the target task never looks.
Reality
Liquid AI shipped a 279.5-million-parameter draft model for LFM2.5-VL-3B on September 24, reporting 2.30x to 3.13x faster decoding on an Apple M5 Max. Image encoding and prompt prefill are unchanged, and the company's tests did not cover quantized deployments.
Reality
NV-Reason-CT reads a 300-to-600-slice study as a single 3D volume instead of a stack of frames. NVIDIA published it as an uncleared research foundation whose cited multireader study measured its chest X-ray predecessor.
Reality
At CIOE 2026 the company showed an infrared vision-language model trained on more than five million of its own thermal samples, built to describe gas plumes and electrical faults on the camera's own chip.
Reality
Damo Academy's RADAR reports a mean AUC of 0.913 across 146 findings and beat 23 of 26 radiologists, and its weights ship under a research-only licence, so commercial deployment still goes through Alibaba.
Reality
Alibaba's DAMO Academy put RADAR's training code and pretrained checkpoints under Apache 2.0 on September 18, a day after the Science paper. The 146 in the coverage counts radiological findings, cancers among them.
Reality
A new synthetic dataset gives Arabic OCR training six fonts and 794.6 million words of perfectly aligned ground truth. Every page is rendered clean, so none of it tells you how a model copes with a bad photocopy.
Reality
A developer ran eight signed radiology studies past Claude, Gemini, Grok and MedGemma, and the sides came back wrong often enough that he moved the laterality check out of the model and into the NIfTI affine.
Reality
Nine local VLM configurations read the same 23 handwritten clinical pages against 221 hand-curated gold values, and accuracy on the fields accepted without human review rose from 75% to 96% once a second model started checking the first.
Reality
A July Nature paper tested whether a general-purpose vision-language model reads a billion-pixel slide better when it searches the way a pathologist does, panning at low magnification before zooming in. The test set was lymph node tissue.
Reality
Ant Ling's documentation samples video at two frames a second, caps clips at 30 seconds and then keeps only 32 frames, leaving the frame budget as the real constraint a GUI agent has to plan around, well short of the million-token context window.
Reality
Alibaba's new 27B model defaults reasoning_effort to xhigh. Simon Willison measured 22,276 reasoning tokens and 21 minutes for one SVG that took two minutes with reasoning off.
Reality
Alibaba scheduled a 27-billion-parameter vision-language model for August 14, alongside an already-published 2.4-trillion-parameter MoE. The smaller file is the consequential one.
Reality
Two releases, two licences. Only the 27B is Apache 2.0, and because just 16 of its 64 layers keep a KV cache, long context costs a quarter of the usual memory.
Reality