In the news
Inkling 2 Mini on vLLM: Day-0 Support for a Looped Model
vLLM · Inferact, Thinking Machines and NVIDIA · Published · 3 min read
In 30 seconds
- What happened
- vLLM added day-0 support for Inkling 2 Mini, a looped multimodal model with 26B total and 3.5B active parameters running on NVIDIA hardware.
- Why it matters
- Engineers deploying or evaluating Inkling 2 Mini need this support to serve the model efficiently with its variable-loop research checkpoint option.
- Watch out
- The looped architecture requires 48 separate KV caches instead of 24, and DSpark speculative decoding requires the default three-pass configuration.
- llm
- speculative
- kernel
- kv cache
- vllm
The patterns behind this
Each one covers how the technique works, when it earns its cost, and where it breaks.
The Agent Architect
One pattern, one tradeoff, one production failure story. A short weekly briefing for people building agentic systems.
Weekly email, one-click unsubscribe. We only use your address to send the briefing.