There are so many awesome people in our local AI community building in the open, so wth, I guess it's my time to show you guys what I'm spending my tokens on lately and how I managed to keep my only DGX Spark burning in A/B testing 24/7 in the last 2 weeks. Many more screenshots to come so fallow the thread.

The Custom Aeon DFlash2 Drafter is coming along great! My retrained NVFP4 version is outpacing even the BF16 version now not just in Tok/s but even acceptance in many categories. Early release dropping for paid Patreon Subscribers then open wide release. patreon.com/cw/AeonForge7/me…

Absolutely great inference atlas, but sir, you do need to improve the UX and make a more intuitive version of it as it is so confusing to browse through it and it really takes time to wrap your mind around it. I've seen the 221 t/s result in the main table for Qwen 27B and wanted to find the recipe and more basic info on it and first I had to scroll through all the 27B recipes that were not sorted by speed and in the same column there were different benchmarks measurements shown and in a total random order. I found the 221 t/s recipe and click on it to be again inundated with lots of info but I still could not see critical info like at what concurrency number, what context size those numbers were obtained and what was the total context pool. Then on the final recipe card (on a small laptop with browser tabs as a sidebar) 90% of the screen is filed with non useful information for 95% of the people and the box with the useful info shows just 5 lines of text and you have to scroll 28 pages of 5 text lines each to go through that info. All that info is amazing and thank you for your huge effort to build this and do all these benchmarks but the most important info for regular users should be on focus and the extra engineering info should be there just for those few that need it as secondary information because just a few users are engineers and will actually read through it.

You now have access to jev models that are better then jev in decisions! 🚀

Many people asked me about the GLi* evaluation tree, and this post is all about it.  Actually, the story starts before GLiNER. In 2022–2023, Chinese research teams and international collaborators explored several complementary approaches. UniMC (2022) framed zero-shot classification as multiple-choice prediction, encoding text and candidate labels together. USM (2023) matched schemas with text through token-level links, unifying entity, relation and event extraction. UniEX (2023) combined span detection, classification and relation extraction within one schema-driven extraction framework. On October 2023, Knowledgator published “As GPT4 but for token classification”, showing prompt-guided token classification for NER, question answering, relation extraction and other tasks. In November 2023, Urchade Zaratiana, and colleagues introduced GLiNER, using a bidirectional encoder to match text spans with entity labels supplied at inference time. Trained on the UniversalNER team’s Pile-NER dataset, it reported strong zero-shot NER results against the evaluated LLM baselines. I was excited about it, subsequently joined its development, and continue to co-maintain the project. In April 2024, Urchade and colleagues introduced GraphER, a related research direction that models entity and relation extraction jointly through graph structure. That June, Mykhailo Shtopko and I introduced GLiNER multi-task, extending the approach to question answering, summarization, and relation extraction. Our team also released GLiClass and bi/poly-encoder variants in 2024. In 2025, Jack Boylan and colleagues introduced GLiREL for zero-shot relation extraction; Robin Armingaud and Romaric Besançon developed GLiDRE. And Urchade and the Fastino team released GLiNER2, combining NER, classification, relation extraction, and structured extraction. In 2026, our team introduced GLiNER-Relex for joint entity and relation extraction. Fastino continued with GLiNER2.5 in August, adding boundary-based extraction, longer context, and constrained joint decoding. And this September, we released GLiFormer. It brings entities, relations, classification, PDF processing, and nested records into one framework with a shared encoder. For hierarchical extraction, it grounds values in the source, groups them into records, and predicts their relationships before assembling JSON. I’m proud of our contribution and grateful to all researchers, contributors, and users helping this ecosystem grow. Looking forward to more exciting contributions from the community.

updated Decision Index 0.1 → 0.2 🎯 better formula, +29 jev-like models, +21 benchmarks AutoJev-27B by @perplexity_ai CTO @denisyarats took the open lead, trailing jev by 0.8 points 🏆 come find the best model at every size, speed and use-case huggingface.co/spaces/multim…