Insights
11,840 TPS on a single GPU
Learn how we optimized NVIDIA Nemotron 3.5 Super VL inference with five vLLM configuration changes, nearly doubling throughput from 6,058 to 11,840 tokens/sec on a single GB300 GPU.
Ara Khan
Insights
DeepSeek wins IMO Gold on 12 cents
Benchmarking eight AI models on IMO 2026 problems in Cline, we found DeepSeek V4 Flash scored 30/42, clearing the gold medal cutoff for just $0.12. See how open-weight models compare with GPT-5.6 Sol, Claude Fable 5, Kimi K3, and more.
Insights
Open-sourcing evals for open-weight agents
Learn how Cline evaluates and improves open weight coding agents using Terminal Bench, with practical heuristics for model performance, token efficiency, reasoning, providers, and eval optimization.
Insights
Recursive Self Improvement for Coding Agents
See how recursive self-improvement helped Cline optimize Kimi K3 and achieve an 88.8% SOTA score on Terminal-Bench 2.1 at a fraction of the cost.
Guides
How to Save Millions by Self-Hosting LLMs
A practical guide to the economics of self-hosting open-weight LLMs. Using Kimi K2.6 and real production traffic from Cline, this post breaks down GPU memory, inference, batching, pricing, and the point at which self-hosting can save millions.
Announcements
A practical guide to hill climbing
We didn't have benchmark numbers, so over a weekend we ran Cline against 89 coding tasks, diagnosed every failure, and shipped fixes that took our score from 47% to 57%. Here's the hill climbing process so you can do it too.
3 Seductive Traps in Agent Building
In building AI agents at Cline, we've discovered that the most dangerous ideas aren't the obviously bad ones, they're the seductive ones that sound brilliant in theory but fail in practice.