Ollama v0.33.0 Fixes KV Cache Bug That Forced Full Reprocessing of 46K-Token Prompts

Ollama released v0.33.0 as a pre-release on August 21, 2025, focused on tighter Claude Desktop integration and fixes to prefill caching that could force full reprocessing of long prompts.

What changed

The release adds direct control over which Ollama models are available inside Claude Desktop, managed from the menu bar. Users can toggle individual local models on or off, and select from available Ollama models within Claude's interface; cloud models only appear when signed in. A new Apps view provides a central place to manage app integrations, including copyable commands.

The most significant technical fix addresses a caching bug affecting agent clients that cancel long prefills. Previously, a cancelled prefill could invalidate restore points that failed to cover what they claimed to, forcing full reprocessing. According to Ollama, on models with recurrent layers this could force a request matching 46,000 of 47,000 tokens to reprocess entirely from zero. The fix makes restore points "trustworthy by construction" — a cancelled prefill now retains every restore point it crossed, so retries resume from the last valid checkpoint instead of restarting.

A related fix disables Claude Code's "tokens left" countdown system message. Ollama had been moving this message to the front of the prompt, which broke KV cache hits on every single request — a change that would have degraded performance for any user relying on prompt caching.

Other changes

The DeepSeek Harness launcher now falls back to npx when a global npm install fails, with added Windows command-shim support. The onboarding flow received updated copy, a macOS header aligned with native traffic-light window controls, and disabled Cmd/Ctrl zoom shortcuts during onboarding to preserve fixed window scaling.

The release also includes an MLX dependency update (PR #17886), a fix for default packaging that had broken due to macOS-specific assumptions leaking into Linux and Windows builds, and a UI fix synchronizing header padding animation with sidebar width animation to stop the Apps header from overlapping macOS traffic-light controls during sidebar transitions.

Full changelog: v0.32.15...v0.33.0-rc2.

What this means

This is a tooling and integration update, not a new model release — no new weights or checkpoints ship in v0.33.0. The KV cache fix is the notable engineering item: any system that silently breaks cache hits on every request has a direct, compounding cost on inference latency and compute spend for users running agentic workflows with Claude Code. Teams running Ollama with recurrent-layer models and long-context agent sessions should update to avoid the full-reprocessing penalty on cancelled prefills. The Claude Desktop menu bar integration signals Ollama continuing to position itself as a local-model companion to hosted assistants like Claude, rather than a standalone competitor.