What changed
Ollama 0.40.15 adds support for MiMo V2.6 and AutoTrust decision models, migrates Kimi Code CLI and improves model discovery for Pi and Claude integrations. It also adds multi-request handling for recurrent and vision GGUF models.
Why it matters
The release broadens which local agent and decision models can run behind Ollama and changes concurrency behavior for recurrent and vision workloads. That can simplify local routing, but model identity, memory pressure and request isolation need checkpoint-specific tests.
What to do
Pin 0.40.15, load one exact supported checkpoint and issue concurrent synthetic requests while invoking the intended CLI integration. Record hashes, prompt-template identity, memory, latency, cross-request leakage, tool output and rollback to 0.40.14.
Caveat
The release notes are maintainer evidence and model support does not reproduce model-card claims. Hardware fit, quantization, license, prompt templates and concurrent-request isolation vary by checkpoint and runtime configuration.