Category 02

AI Model Serving & Inference

Model hosting, inference engines, optimization and deployment.

Standout picks across editions

The current edition's pick, highlighted, plus earlier picks in this category from our backfilled archive — sourced from credible outlets (Hacker News, GitHub, TechCrunch and more) and checked against each product's official site.

This edition's pick · Nov 16–30, 2026
ST

Sarvam Token Factory

India-hosted inference via an OpenAI-compatible API

India-hosted model inference, OpenAI-compatible API.

Why it matters

Opened for general use on Nov 16 — a sovereign, in-region inference option with an OpenAI-compatible surface, lowering the switching cost for teams with data-residency needs.

Oct 1–15, 2026 edition
I

Indra

1M-param OCR model, runs on CPU

1M-param OCR model, CPU-fast, near billion-param accuracy.

Indra, from Datalab, is an OCR model that turns PDFs and images into markdown, text or JSON with bounding boxes. At roughly 1 million parameters it runs entirely on CPU — no GPU or PyTorch — yet reaches about 15× Tesseract's speed and scores 66.7 on OmniParseBench versus PaddleOCR's 58.4. The code is Apache-2.0, with a modified OpenRAIL-M license on the model weights for research and personal use.

Why it mattered

A 1M-parameter OCR model that runs on CPU ~15× faster than Tesseract while approaching billion-parameter accuracy — efficient inference from Datalab.

Evidence dated 2026-10-13github ↗official ↗
Sep 16–30, 2026 edition
SD

Strands Decider

Fast decision model for agents

Small, fast decision model with calibrated confidence.

Why it mattered

A small, fast 'system-one' decision model with calibrated confidence for agentic workflows — picking or rating options faster than an LLM.

Evidence dated 2026-09-29github ↗
Researched · pending approvalVisit site ↗

How picks are selected

Candidates are scored on market traction, recent momentum, product innovation, community interest, practical usefulness and source credibility. Only researched, fact-checked products are published — if no candidate meets the evidence threshold, we show no pick rather than forcing a weak selection.