SShortSingh.
Back to feed

9B model tops DeepSeek V4 Pro on math benchmark; caveats apply

0
·1 views

A TikTok clip claimed a 9-billion-parameter open model beats DeepSeek V4 Pro on math tests, and a review of the Hugging Face listing confirms the cited figures exist. The model is a fine-tune of Qwen3.5-9B trained on roughly 250,000 answers generated by DeepSeek-V4-Pro, released under an Apache-2.0 license. On GSM8K it scores 94.50% versus 92.60% for DeepSeek-V4-Pro, though the test conditions differ from those used for other comparisons. The reviewer notes the benchmarks are self-reported, GSM8K is public and may overlap with training data, and the model lists no Thai among its supported languages. Its uncensored community variant also refuses far fewer requests, so output safeguards are needed in any product use.

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)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Goal Engineering: ออกแบบเป้าหมายเอเจนต์ให้คนสอนได้

เมื่อค่ำวันที่ 14 พฤศจิกายน 2026 เวลา 23:05 น. ตามเวลาไทย แฮร์ริสัน เชส (Harrison Chase) จาก LangChain เผยแพร่บทความบน X ชื่อ "Goal Engineering: How to Design Goals" เขียนร่วมกับ มิฮีร์ การิเมลลา (Mihir Garimella) และอันชุล กุปตะ (Anshul Gupta) จาก Actively AI วันเดียวกันมีฉบับบนบล็อกของ LangChain ด้วย โพสต์นำบอกว่า เมื่อเอเจนต์เริ่มทำงานตามเป้าหมาย การออกแบบเป้าหมายให้ถูกทางจึงสำคัญ ผู้เขียนทั้งสามเสนอว่า ขอบเขตของเอเจนต์ควรออกแบบให้คนให้ฟีดแบ็กคุณภาพสูงได้ง่าย ความรู้ที่ยากที่สุดไม่ได้อยู่ในเอกสารหรือเอพีไอ แต่อยู่ในหัวผู้เชี่ยวชาญ เคสหลักคือ Actively AI ที่ย้ายจากเอเจนต์ต่อหนึ่งบัญชีลูกค้า

0
ProgrammingDEV Community ·

Base64, Hex, Base32 and ASCII85 Compared: Trade-offs in Size and Readability

Base64 is the most common way to encode binary data as text, but alternatives such as hexadecimal, Base32 and ASCII85 each exist to address specific constraints. Hex doubles data size but is easy for people to read, while Base32 adds about 60% overhead to avoid confusion from letter case and similar-looking characters. ASCII85 is the most space-efficient of the four at about 25% overhead and is used in PostScript and PDF files. Base64's dominance comes from carrying 6 bits per printable character, with its remaining issues handled by URL-safe variants. The article also warns that hex strings can be decoded as Base64 without error, producing incorrect bytes.

0
ProgrammingDEV Community ·

Price per GB of Unified Memory Ranks Strix Halo Mini PCs for Local AI

A 128GB mini PC at $3,500 and another at $4,349 show that sticker prices alone reveal little about value for local AI work. Because the RAM is soldered and cannot be upgraded, the author argues the fairer measure is price per gigabyte of unified memory, and a ranking of 15 Strix Halo configurations uses that metric. The GPU shares system RAM, so the usable GPU memory determines which models run; for example, a 64GB machine gives about 48GB to the GPU, while a 128GB machine gives about 96GB. The cheapest per-GB option in the top ten is a 96GB Bosgame M5 at $24.99/GB, but it cannot run a 120B mixture-of-experts model, making the Framework Desktop 128GB at $26.95/GB the cheapest route to 96GB of GPU memory. Prices vary between Amazon and official stores and have shifted over short periods, so the author advises checking both channels and dates before buying.

0
ProgrammingDEV Community ·

React Native and Expo apps still expose their JavaScript bundle inside the installed app

Apps built with React Native or Expo ship their JavaScript or TypeScript logic inside the native binary, whether as a plain bundle or as Hermes bytecode. That code can be extracted from the APK, IPA, or an over-the-air update by anyone who looks, so being a mobile app does not protect it. The article notes that Hermes raises the effort needed to read the code but does not encrypt it, and that OTA update payloads can be an easier source to copy. It advises keeping entitlements, secret keys, and authoritative pricing on the server, and applying real protection such as obfuscation to any sensitive logic that must run locally. It also warns against shipping public source maps or unprotected debug bundles.