DEEPSEEK
28 items
25 posts, 3 tools
A first-hand visit to DeepSeek HQ reveals something more interesting than benchmark scores: a 300-person company that treats AI as infrastructure, not eschatology - and what that means for API pricing everywhere.
Reasonix builds a DeepSeek coding agent loop around prefix caching. When cache discipline keeps long sessions cheap, and when it fights the harness.
Open-source (MIT) terminal coding agent, renamed Codewhale, with Plan, Work, and Operate modes, approval postures, /undo and /restore, receipts, MCP servers, sub-agents, and a local runtime HTTP API. Provider-neutral: 40+ built-in provider routes plus local models, no longer DeepSeek by default.
DeepSeek V4 brings near-frontier coding at a much lower token price. Where cheap reasoning fits in a coding agent stack, from code review to repo mapping.
DeepSeek V4 developer guide: Flash vs Pro, pricing, OpenAI SDK setup with thinking mode and tool calling, and when to pick V4 over Claude or GPT.
GRPO and PPO explained for developers with no RL background: how DeepSeek R1 was trained, why GRPO drops the value model, and a minimal GRPO run with TRL.
DeepSeek's reasoning-first model built for agents. First model to integrate thinking directly into tool use. Ships alongside V3.2-Speciale, which rivals GPT-5 and Gemini 3.0 Pro.
DeepSeek R1 and V3 developer guide: use the MIT-licensed models through the API, run them locally with Ollama and decide when they beat closed-source models.

Get Smarter About AI Dev
New tutorials, open-source projects, and deep dives on coding agents - delivered weekly.