DeepSeek R1

What It Is

DeepSeek R1 is an open-weight reasoning model that uses Mixture-of-Experts (685B total, 37B active per query) and reinforcement learning to develop chain-of-thought reasoning without supervised fine-tuning on reasoning examples. The R1-0528 update pushed AIME 2025 accuracy from 70% to 87.5% and LiveCodeBench from 63.5% to 73.3%. Distilled variants (8B, 32B, 70B) make reasoning accessible on consumer hardware.

Why It Matters

DeepSeek R1 is Emerging because it proved something important: reasoning capability can emerge from RL alone, and open-weight models can compete with frontier proprietary ones. The MIT license means you can download, modify, and deploy it however you want. The 8B distilled variant runs on consumer GPUs while retaining substantial reasoning capability.

The complications are significant. Cisco testing found a 100% jailbreak success rate — unprecedented among frontier models. CCP censorship is baked in (the model refuses or deflects on politically sensitive topics). Seven-plus countries and dozens of U.S. agencies have restricted its use. These aren't dealbreakers for every use case, but they constrain where you can deploy it.

Key Developments

• R1-0528: AIME 2025 accuracy jumped to 87.5%. GPQA-Diamond from 71.5% to 81.0%. Deeper chain-of-thought averaging 23K thinking tokens per query.
• Jan 2026: Updated research paper expanded from 22 to 86 pages documenting training methodology.
• 2026: 8B distilled variant released (R1-0528-Qwen3-8B) for consumer hardware deployment.
• Ongoing: 7+ country bans and dozens of U.S. agency restrictions on government devices.

What to Watch

The March 2026 reports of a new DeepSeek model trained on NVIDIA's most advanced AI chip are the signal. If the next iteration closes the safety gap while maintaining open weights and MIT licensing, it could move to Promising despite the geopolitical concerns. Watch for the security vulnerability response — whether DeepSeek patches the jailbreak vectors that Cisco identified.