NVIDIA Nemotron 3.5 Lightning context management with Entroly and Ollama
Run the model locally with Ollama while Entroly selects evidence under a budget, keeps omitted originals recoverable, emits Context Receipts, and optionally checks the response against supplied evidence.
nemotron-3.5-lightning through its existing local-model discovery and OpenAI-compatible proxy path. This is real model-neutral compatibility, not a claim that Entroly modifies NVIDIA's model or Ollama runtime.Verified model and integration boundary
| Surface | Verified behavior |
|---|---|
| Ollama model | nemotron-3.5-lightning; Ollama describes it as a 30B mixture-of-experts model with 3B active parameters and tool support. |
| Context | Ollama lists latest/30b at 1M and 30b-mlx at 256K. Entroly can inspect the local tag's reported context length. |
| Entroly path | Loopback-only Ollama discovery plus an OpenAI-compatible upstream at http://127.0.0.1:11434. |
| Privacy | Entroly sends no analytics by default. Local Ollama inference can keep prompts on-device; tool calls and separately configured services may still use the network. |
Setup
ollama pull nemotron-3.5-lightning pip install -U entroly python -m entroly.models discover ollama --inspect-ollama-context # Configure this environment variable in your shell: ENTROLY_OPENAI_BASE=http://127.0.0.1:11434 entroly proxy
Point an OpenAI-compatible client at Entroly's local endpoint, normally http://127.0.0.1:9377/v1, and select nemotron-3.5-lightning. Client-specific environment-variable syntax differs across PowerShell, Command Prompt, Bash, and application settings.
Why context control still matters with a long window
A large advertised window is capacity, not a requirement to resend every file, log, conversation, and retrieval result. Entroly can select answer-relevant evidence, preserve exact omitted content through local recovery handles, keep stable prefix bytes cache-friendly, and record the decision without claiming that fewer tokens automatically produce a better answer.