DeepSeek
A reasoning-first model that surprised everyone who wrote it off too early. Cost-efficient, genuinely open, and strong on hard problems — one of the better quiet arrivals of the past year.
Scoring · in review
Every score is hand-assigned across five axes — utility, privacy, speed, cost, transparency. This one isn't published yet because we're still using it long enough to have an honest opinion. Slow scoring is the point.
We don’t earn anything when you click this.
Who it's for
Move faster with an AI co-pilot
Offload the repetitive, automate the tedious, and focus on the creative parts of your work that actually need you.
Learn by doing, not by watching
Ask it anything, get back an explanation calibrated exactly to your level. The best tutor you've ever had doesn't charge by the hour.
The Assessment
- ✓ Reasoning
- ✓ Open-source
- ✓ Efficient
- × Refuses politically sensitive prompts with no transparency about what triggers refusals — can be unpredictable.
- × API stability has been variable; not yet suitable as a primary production dependency for critical workloads.
- × Data handling and residency policies are less transparent than US-based alternatives.
Related concepts
How modern models pretend to be huge while doing the work of something smaller.
ConceptQuantizationWhy a 70B-parameter model can run on your laptop — and the quality you trade for it.
ConceptReasoning ModelsWhat changed when models started thinking before they answered.
ConceptSynthetic DataModels training on text other models wrote — and why this isn't always bad.
ConceptHow AI Models Are TrainedFrom random noise to a model that can reason — the actual pipeline
What DeepSeek can’t do
It's fast and cheap. You decide whether fast and cheap is what the question needs.
DeepSeek has remarkable performance at a fraction of frontier cost. The question of whether to use it — whether speed is what matters here, or whether the problem deserves a more careful model — is still yours.
Read the essay →