We need to produce JSON with fields: seo_title, summary, key_entities (companies, people, locations), expert_faq (question, answer), internal_link_anchor, primary_keyword, secondary_keywords (list), content_type.

We must rewrite the news as a polished 180-250 word summary, split into 3-4 paragraphs, each 45-65 words. Must bold a few key details with double asterisks, 3-5 bolded terms total. Must not use banned vocabulary etc. Must be valid JSON.

We need to extract info from source: FUTO Swipe, a swipe typing model. FUTO Keyboard app offline Android. Dataset collection August 2024, >1 million swipes, dataset released March 2025 under MIT license on Hugging Face. Models: Encoder (layout-agnostic, language-agnostic), Context LM (small LM per language), Decoder (language-specific layout-specific). With beam width 300 achieve top-4 fail rate ~4% on test set; ignoring OOV error below 1%. Footprint: encoder 635,140 params, decoder 304,155 extra, Context LM 1.5M (1.1M embeddings) => active parameters 1,364,271, total 2,494,767. Runs on low-end devices in milliseconds. Environmental cost low: only needed 1 workstation GPU. C++ Library swipe-library for inference, beam search. License: FUTO Model License for models, GPL for library. Working on a paper.

We need to bold key details: maybe FUTO Swipe, 1 million swipes, 2.5 million parameters, MIT license, GPL. Choose 3-5.

We must not use banned words: list includes additionally, align with, crucial, delve, emphasizing, enduring, enhance, fostering, garner, highlight, interplay, intricate, intricacies, key, landscape, pivotal, showcase, tapestry, testament, underscore, valuable, vibrant, nestled, groundbreaking, renowned, breathtaking, must-visit, stunning, profound, game-changer, robust, paradigm. Also banned filler phrases. Must avoid them.

We must end with concrete statement, not forward-looking platitude.

We need to count words 180-250. Let's craft ~210 words.

Paragraph 1: ~55 words. Paragraph 2: ~5