What Happened
Ollama 0.40.11 adds speech generation on macOS: the CLI can play replies as they are generated, and applications can stream speech through the OpenAI-compatible /v1/audio/speech endpoint. It also adds d3 decision models in 2B, 4B, 9B and 27B sizes for classification, yes/no answers and probability scoring. [1]
Ollama 0.40.9 extends Flint to Linux and Windows. It also makes streaming OpenAI Responses and Anthropic Messages requests report generation failures rather than silently ending after partial output. Cloud-model context lengths and capabilities are now passed to coding tools through ollama launch. [2]
LiteLLM 1.103.8, 1.104.6, 1.105.4 and 1.106.3 use a new signing key for Docker images and Helm charts; the previous key will not verify those artifacts. These releases also remove the eMudhra root CA from trusted certificate bundles. Version 1.103.8 fixes OpenTelemetry callback initialization with OpenTelemetry 1.44 and newer; 1.104.6 includes fixes across Azure AI Foundry requests, streaming, tool calls, token counts, caching, billing and usage reporting. [3][4][5][6]
A Streamlit nightly build is identified, but no release details are available to establish feature or breaking-change claims. [7]
Why It Matters to Businesses
The most immediate upgrade risk is deployment, not model behavior: LiteLLM image or chart verification will fail wherever CI or admission policies pin the old signing key. Ollama’s streaming fix also changes a failure mode applications may have mistaken for a successful, shortened answer. [2][3]
Kimbodo Engineering Perspective
We would treat these as separate changes to the artifact trust chain, streaming contract and application capabilities. A signed release that cannot pass admission is an outage risk; a stream that ends without an explicit error is a correctness risk. Speech and larger decision models are optional capabilities whose latency, hardware needs and operating costs should be measured before rollout.
How We Would Implement It
- Pin target versions and update LiteLLM verification policies only after obtaining the new cosign public key through a trusted channel. Check its SHA-256 against e3fe3e08c44b16407fe0b37c43b4f8b603b3a0f4ee74c29e647bc022d762c2c0, then test image and chart admission in staging. [3]
- Add streaming tests that force a generation failure after partial output and confirm clients surface an error rather than treating the response as complete. Recheck coding-tool context settings for Ollama cloud models. [2]
- Pilot speech on macOS behind an application feature flag; measure first-audio latency, concurrent stream capacity and output handling before enabling it for users. Benchmark d3 sizes against the specific classification tasks they would replace. [1]
- Run LiteLLM regression tests for the provider integrations, telemetry and usage reports relevant to the deployed version. [5][6]
Risks, Costs and Security
Changing a pinned signing key requires controlled verification, not simply disabling signature checks. The CA removal may also affect connections that depend on the removed root; test those integrations before promotion. Speech introduces audio-output handling and capacity costs, while larger models increase resource requirements. Keep the Streamlit nightly out of change plans until its release details can be verified. [3][7]
Where Kimbodo Comes In
Kimbodo builds and operates this in production for businesses — see our AI Application Development practice, or Estimate My AI Application.
