Table of Contents
Open-weight LLMs ship faster than enterprise software. New flagship models roughly quarterly.
TL;DR
Major open-weight releases: Meta (Llama) ~6 months, Mistral ~3 months, Alibaba (Qwen) ~3 months, DeepSeek ~6 months. Plan for a model upgrade every quarter.
Release cadence
- Llama: 3.0 (Apr 2024), 3.1 (Jul 2024), 3.2 multimodal (Sep 2024), 3.3 (Dec 2024)
- Qwen: 2.0 (Jun 2024), 2.5 (Sep 2024), 2.5 Coder (Oct 2024)
- Mistral: 7B v0.1, v0.2, v0.3, Small, Large, Nemo, Codestral
- DeepSeek: V2 (May 2024), V2.5 (Sep 2024), V3 (Dec 2024), R1 (Jan 2025)
Strategy
- Pin commit SHAs in production
- Quarterly eval against latest releases
- Migrate when delta >3% on your eval set
- Don't auto-upgrade; validate first
Verdict
Plan for quarterly model evaluations. Don't over-tune to a specific model — the next one is 3 months away.
Bottom line
Open-weight moves fast. Build the eval pipeline early. See eval pipeline.
