DeepSeek
Open-weights LLM family from a Chinese AI lab. One of the strongest open-source models available as of early 2026 — frequently cited alongside Qwen as having surpassed Meta’s Llama in capability. The R1 reasoning line is the wiki’s default recommendation for “free local-equivalent reasoning model.”
R1-0528 (May 2025 refresh)
The latest R1 release at the time of WorldofAI’s benchmark:
- Params: 671B total, 37B active (Mixture-of-Experts)
- Context: 136K
- License: MIT — fully unrestricted, no Commons Clause
- Pricing: 5/M output via DeepSeek API; free tier on OpenRouter
- MMLU: competitive with OpenAI O3-high and GPT-4 mini
- Demoed: 107s chain-of-thought on multi-step unit-conversion math (correct); full animated SaaS landing page in one shot
R1-0528 is now the default open-weights backbone for several tools in the wiki: Codename Goose (recommended free model), Stagehand (R1-distilled Qwen 32B is the fastest+cheapest in Stagehand’s eval), DeerFlow (named explicitly as a recommended model), AgenticSeek.
Why It Matters
DeepSeek demonstrated that open-source models from non-US labs could match or exceed models from established US AI labs at a fraction of the training cost. Its release caused significant attention in the AI community regarding compute efficiency.
Use in Local AI
Available in GGUF format via llama.cpp and Ollama. Commonly referenced in quantization naming: DeepSeek-R1-Q4_K_M follows the standard llama.cpp quantization naming convention.
Compared to Other Open-Source Models
Among the current leaders alongside Qwen. Both are considered stronger than Llama for most tasks as of early 2026.
API Access
Also available via DeepSeek’s own API and through OpenRouter for cloud inference at low cost.
See Also
- Qwen — comparable Chinese lab model
- Llama — Meta’s model; now behind DeepSeek in capability
- llama.cpp — inference engine for running locally
- Open-Source Model Integration — using open-source models with Claude Code
- Matthew Berman — source
- Source: Every AI Model Explained