# Open-Source and Open-Weight Models

Open models (Llama, Mistral, DeepSeek) have reshaped the frontier. This sub-article expands [Section 1 of Recent Topics](../recent.md) and [Hardware & Software](../hardware-software.md).

## The Open Model Race
- Meta released **Llama**; Mistral and DeepSeek released competitive open-weight models.
- China's Qwen, GLM (Z.ai), and Kimi (Moonshot) now release open weights rivaling US frontier models.
- Distillation produces small models (Llama 3.2, DeepSeek-R1, Qwen3.8 27B) usable on laptops.
- See [Frontier Labs](frontier-labs.md) and [Frontier LLMs](frontier-llms.md) for the current model landscape.

## Benefits
- Reproducibility, research access, and innovation.
- Lower cost and broader deployment, especially for institutions without frontier API budgets.

## Risks
- Malicious actors can fine-tune open models.
- Safety mitigations (refusals, guardrails) are weaker than in hosted APIs.
- Regulation and monitoring are harder when weights are public.

## Governance Debate
- Some call for licensing or compute thresholds; others argue openness aids safety research.
- The open-vs-closed debate is central to AI policy.

## Why It Matters
The open-model race affects who can build, audit, and misuse AI.

## Sources
- Meta. *Llama release notes.*
- DeepSeek. *Technical reports.*
- Mistral AI. *Model documentation.*

*Back to [Recent Topics](../recent.md) · See [Frontier AI](frontier-ai.md).*