MN Uncensored v0.3.1
Cost-safety notice: Keep all models hard-stopped until a Modal Workspace budget is configured. The v0.3.1 CLI still contains the old persistent-warm
mn startbehavior. Use the fail-safe code onmainand read the 2026-07-16 incident report. A safety runtime release is intentionally pending until the budget is confirmed.
Released 2026-07-16.
Fixed
- Replaced five-second cold-start polling with exponential health-check
backoff. - Health probes now wait 30, 60, 120, and 240 seconds, capped at five minutes.
- Prevented long 35B starts from accumulating a queue of redundant Modal GPU
invocations.
Runtime status
- Catalog:
mn/god,mn/code, andmn/fast. - Lifecycle: independent request-triggered start and ten-minute idle shutdown.
- Context: 131,072 tokens on every route.
- Output ceiling: 16,384 tokens.
- Release verification: streaming completion and forced tool call on every
model.