Search

LOAD_ONNX_MODEL

This reference describes how to load ONNX embedding and rerank models into a database from a file or BLOB, with syntax, parameters, and examples. It also lists naming and size limits, metadata requirements, external-data handling, and memory constraints.

Nex (@NexEcosystem) / X

NexRT is a low-latency decode engine for Nex-N2.5 aimed at single-request inference on Nex-N2.5-Pro, delivering 1,000+ tokens/s with 8K context on 8×H100 using DFlash, supporting standard decoding, MTP, and DFlash, with context parallelism for very long contexts (up to 256K).

Introducing Omnii-1: a leap toward general biological intelligence — Radical Numerics

Omnii-1 introduces a general, multimodal genome language model intended to understand, predict, and engineer biology across systems. It claims transferability across tasks, reports pretraining on 28 modalities and trillions of tokens, and presents eight applications built on the shared foundation.

krea-2/README.md at db3984fbc6e13b34c0064990fc2d95ac64d00058 · krea-ai/krea-2 · GitHub

Two models work together: train LoRAs on RAW and apply them on Turbo for inference; LoRAs trained on RAW perform well on Turbo.

Anthropic's initial public offering shows a margin of 85.4% compared to Z.ai and MiniMax - News and statistics - IndexBox

Translated title. Original: L'introduction en Bourse d'Anthropic affiche une marge de 85,4 % face à Z.ai et MiniMax - Actualités et statistiques - IndexBox

Anthropic's IPO prospectus shows a large profitability gap versus Chinese rivals Z.ai (Zhipu AI) and MiniMax, highlighting monetization challenges for Chinese AI developers; Anthropic reports 85.4% gross margin in H1 2026, compared with 26.4% for Z.ai and 17.9% for MiniMax, according to SEC filings and market analysis. This underscores competitive dynamics between US and Chinese AI firms in revenue and margins.

Q8_0/Swift-1.5-Qwen3.8-Flash-Next-Q8_0-00002-of-00005.gguf · ukisai/Swift-1.5-Qwen3.8-Flash-Next-GGUF at main

This document provides step-by-step instructions to install and run the ukisai/Swift-1.5-Qwen3.8-Flash-Next-GGUF model using llama.cpp, Docker, or vLLM, including server setup, inference commands, and cross-platform installation paths (macOS, Linux, Windows).

BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model

Tether AI Research presents BotPsy, a 0.8-billion-parameter model for on-device tool use and personal assistant tasks. The document reports strong benchmark results for its size, while noting limits in instruction following, language coverage, and high-stakes reliability.

Release b11684 · ggml-org/llama.cpp · GitHub

The document documents test updates and baseline handling for a fusion baseline in a code repo, including re-recording baselines, regression reporting when matches fail, and tweaks to Mellum MoE routing and GGUF-based models; it also notes how disclosures and assistant credits are managed.

Hassabis Trims AGI Window to Two-to-Three Years in WSJ Profile | AI Weekly

Demis Hassabis projects AGI within two to three years, shortening his prior four-year horizon; the Wall Street Journal frames this as their wording. The piece cites Hassabis's Einstein test for progress and notes Isomorphic Labs and its planned funding round, illustrating AI's biomedical and scientific potential.

GitHub - oguzeldereli/ttf-home: Ternary language models on a home computer: run, train, convert, fine-tune, export to GGUF · GitHub

ttf-home is a command-line toolkit for training and running ternary language models, converting and fine-tuning models, and exporting them as GGUF files for llama.cpp. Its instructions include setup, serving, GPU requirements, model sizes, and performance results.