v0.4.1 v0.5.0

Name: convex-backend
Author: convex

convex/convex-backendNov 10, 2025 → Nov 14, 20254 days between releases

Generating AI comparison

Analyzing what changed between v0.4.1 and v0.5.0. This page updates automatically — usually within a minute.

The Two Releases

v0.4.1

Nov 10, 2025· 1 year ago

Fixes a critical memory bug on Blackwell GPUs where the RoPE (Rotary Positional Embedding) cache incorrectly expanded to an excessive number of positions, causing memory blow-ups.

Full release notes

v0.5.0

Nov 14, 2025· 11 months ago

Introduces DSQ (DeepSeek Quantization) to reduce VRAM/RAM usage for DeepSeek-OCR and PaddleOCR-VL models by patching linear layers while keeping embeddings in float. Supports three precision tiers (Q4_K, Q6_K, Q8_0) and parallel loading to improve cold-start times.

Full release notes