Compressed 27B AI model outperforms full version on function calling benchmark

·1 views

Underdog Research released a 7.89 GB compressed version of the Qwen3.8-27B AI model called Saluki 27B. The compressed model scored 88 out of 120 on the team's proprietary function calling test suite in October 2026, while the original 54 GB model scored 84. Saluki uses a mixed 2-bit quantization technique and runs on standard llama.cpp without special builds. The creators note that while Saluki excels in single and multi-choice function selection, its mathematical performance declined by 15-18%.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

New AI app Re:Set aims to help users reduce screen time and read offline.

A team has developed Re:Set, an open-source AI tool designed to combat screen addiction. The app creates personalized offline reading rituals using Google's Gemma 2 model. It features a system for local venues to host verified silent reading sessions. The project was created for the Hacktoberfest 2026 Open-Source AI Challenge.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

DEV Community Article Details Shipping Cutoff Rules for Warehouses

A DEV Community article presents a function designed to calculate whether same-day shipping is possible from a warehouse. The function checks local time against cutoffs, considers holidays, blackout carriers, weekends, and VIP service extensions. It is explicitly labeled as a teaching example, not code from a live production system. The article includes a table of twelve test cases to verify the function's behavior under different time zones and conditions.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

WANDER App Uses Local AI to Generate Personalized Outdoor Missions

The WANDER app is an offline-first AI companion designed to encourage users to explore their surroundings. It uses Google's Gemma 3 4B model running locally via Ollama to generate three personalized outdoor missions based on user mood, available time, and weather. The app operates with a privacy-first approach, storing user journals and activity history locally on the device using SQLite. After selecting a mission, users are prompted to put their phones away and engage with their environment. The project was submitted for the Hacktoberfest Open-Source AI Challenge and is currently in early development stages.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Algorithm solves 'celebrity problem' efficiently with linear time

The celebrity problem asks to identify a person in a group whom everyone knows but who knows no one else. It is typically represented by an n x n matrix where entries indicate whether one person knows another. A naive brute-force solution requires O(n²) time to check all relationships. However, a more efficient algorithm can find a potential candidate in O(n) time using a single pass, then verify the candidate with a second pass. This optimized approach uses constant extra memory, improving on earlier stack-based linear-time methods.

0 comments Read more at DEV Community