We nearly doubled throughput for Nemotron 3.5 Super VL on a single GB300 GPU, from 6,058 → 11,840 tok/s, with five configuration changes in vLLM.
These let the model draft tokens ahead, keep the GPU from ever sitting idle, and store the model’s memory at half precision.
We wrote more about the tricks we used and the future of local models and agentic inference here:
cline.bot/blog/11-840-tps-on…
Kimi K3.1 Pro nearly matches Opus 5.5 performance on Terminal-Bench 4.0 at *half* the cost.
Moonshot also removed cyber data from mid-training and RL to deliberately reduce performance on exploit tasks by ~10%, while everything else, including coding, kept improving.
Now available on ClinePass, our $9.99/month subscription plan for ~5x discounted access to open weights models.
cline.bot/cline-pass
There’s new free models available in Cline:
- Reflection Beam (501B)
- Nemotron 3.5 Super VL (120B)
- Poolside Laguna S 2.1 (118B)
- Nemotron 3.5 Lightning (32B)
- Solar Mini 4 (35B)
6 months ago these Terminal-Bench scores were frontier, open weights makes this more accessible.
Try Cline with:
npm i -g cline
Or install our new desktop app for mac, windows, and linux:
cline.bot/desktop
Cline retweeted
154 DAYS LATER....I Finally Dropped a New YouTube Video🙈
A little late to the party on X, but I just HAD to share this here because I'm so overwhelmed!!
This time, I tested the new @deepseek_ai V4.1 Flash 1017 and used it to build a fully interactive 3D city with @cline.
After being away from YouTube for so long, it feels so good to finally be back!! 🥹
Full video below ↓
We’re about to see a breakout in local model capabilities.
Flint scores 54.9% on Terminal-Bench 2.1 with ~2B active parameters. That’s ahead of every other small MoE and two of the three dense ~30B models.
It only needs about 24GB and runs at 125+ tok/s on a MacBook Pro.
Run it in Cline now:
1) Install Cline:
npm i -g cline
Or use our desktop app:
cline.bot/desktop
2) Configure local models through Ollama:
docs.cline.bot/running-model…
3) Install Flint:
huggingface.co/RecursiveSI/f…









