4 min read
Krea 2 shrinks to 12GB so consumer…
Krea AI open-sourced Krea 2, a 12.9B-parameter diffusion transformer trained from scratch. Turbo runs in 8 steps on GGUF quants that fit 12GB VRAM. Raw is the fine-tune base; Turbo is what you ship.
Blog 385 posts
Notes from the trenches on AI engineering, LLM apps, and the full-stack work that holds it all together.
4 min read
Mistral OCR 4 returns a document map,…
Mistral OCR 4 adds paragraph-level bounding boxes, 13 block types, and inline confidence scores across 170 languages. At $4 per 1,000 pages it is built for RAG chunking, agent grounding, and self-hosted document pipelines.
5 min read
OpenAI's Jalapeño chip targets inference costs, not…
Jalapeño is OpenAI's first custom Intelligence Processor, co-built with Broadcom in nine months for LLM inference. Early benchmarks show 1.5x to 3.6x better latency per watt than today's GPU racks, with volume deployment starting late 2026.
4 min read
Baidu Unlimited-OCR reads 40+ pages in one…
A 3B MoE model with Reference Sliding Window Attention parses long PDFs in a single forward pass. Here is when it beats page-by-page OCR pipelines for RAG and document automation.
4 min read
Builder.io Clips is the screen recorder agents…
Clips is a free open-source Loom alternative where share links expose agent-readable transcripts and metadata. Built on the Agent-Native framework so UI and agents share the same actions.
4 min read
Princeton i1-3B beats FLUX.1 Dev on public…
A 3B text-to-image model from Princeton matches leading models at 1024px using only public training data. Weights, code, data pipelines, and recipes are all open.
5 min read
Sakana Fugu turns model routing into one…
Fugu is a learned orchestrator that picks frontier models per step, swaps providers when export controls bite, and ships as a single API call. Here is how I would wire it into a production agent stack.
4 min read
Self-predicting latents need exponentially less training data…
A new theory proves latent prediction recovers hierarchical structure with constant samples while token-level SSL needs exponential data. Here is what that means for JEPA, data2vec, and your pretraining budget.