Dharma Field Extractor: any document in, schema-exact JSON out

A 2B LoRA you self-host on one L40S — plus Gemini drops an embeddings preview, adds one-line voice remixing, and DeepSeek's fastest Flash hits the Gateway.

Nowline NOV 8 11:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Any document in, schema-exact JSON out

    Dharma-AI shipped Field Extractor, a 2B LoRA on its open DharmaOCR-LITE that turns a document image into structured JSON — either auto-detected key/values or a schema you supply, where missing fields return null instead of invented ones. It's self-hosted on a single L40S (~$2.24/hr), which Dharma puts at 6–147x cheaper than frontier APIs; on its own 1,638-case benchmark it reports 0.834 vs 0.816 for Claude Opus 5 and 0.741 for GPT-5.6. Point it at invoices, forms or medical records and the output maps straight to your DB — no glue layer.

  • Gemini embeddings: preview's gone, swap one string

    As of Nov 6, calls to gemini-embedding-2-preview now route to gemini-embedding-2 at standard rates. Same API schema, same pricing, same embedding space — so it's a one-line model-string change with no need to re-index your existing vectors. Fix it before the preview alias drops out of your pipeline.

  • Remix any Gemini voice with one sentence

    Gemini 3.8 Flash TTS can now reshape a prebuilt or designed voice from a natural-language instruction: POST to /v1beta/voices with gemini-3.8-flash-tts to mint a reusable voice_ ID, then synthesize with 3.8 Flash or Flash-Lite TTS. Tune a narrator or character's tone without recording a session or cloning a real person.

  • DeepSeek's fastest Flash reaches Vercel's Gateway

    Nov 7: DeepSeek V4.1 Flash 1017 'Ultrafast' — the quicker checkpoint of the Flash model already on the gateway — is now callable through Vercel AI Gateway on a single key, with no separate DeepSeek account, plus spend tracking and request traces. Point latency-sensitive calls at it without touching your auth setup.