Dharma's open 2B model turns any document into JSON for $2/hr
It edges Claude Opus 5 on structured extraction — plus IBM's open deck-builder, a 91-class CWE eval, and a new contract for shipping Transformers models.

Copy markdown
A 2B model that out-extracts Opus 5
Dharma-AI's Field Extractor (built on DharmaOCR-LITE v1.1, a LoRA trained on 43,384 examples) takes a document image and returns structured JSON — open-ended or schema-directed. On a 546-doc, 1,638-case benchmark it scores 0.834, edging Claude Opus 5 (0.816) and well past GPT-5.6 Luna (0.741), and self-hosts on a single L40S at $2.24/hr.
IBM's Palette builds slides in your design language
ibm-research open-sourced Palette, a 20B LoRA-tuned model with split designer and coder roles, trained on ~1,000 decks (~10,000 slides), that generates presentations matching your own design system. It ranks first on Economics and Education in PresentBench and trails NotebookLM by just 1.1 points on the three-domain mean (59.7 vs 60.8). Weekend build: a brand-accurate deck generator you host yourself.
CWE-bench v2: stress-test your coding model on real vulns
collinear-ai shipped a public 100-task set spanning 91 CWEs so you can benchmark your own agents on security. Across 11 models, Gemini 4 Argon leads at 67.8% pass@1 / 81.0% pass@4 with Claude Opus 5.5 right behind (67.3% / 78.0%); DeepSeek V4.1 Flash reaches 57.0% pass@1 at $0.04 a rollout.
A packaging contract to stop silent Transformers regressions
Hugging Face proposed transformers-mic, a standard that ships each model with a shrunk fixture, usage path and baselines (e.g. UMT5 at 66 MB vs a 1.2 GB checkpoint; Llama 4 Scout shrunk from 108.6B to 44.7M params). It already caught a two-month UMT5 SDPA regression — a safety net if you integrate new models into transformers.