Pull Request Explorer

Summary

Prevent Metal GPU timeouts during model load

Ensures reliable and faster model loading on macOS by eliminating GPU command buffer stalls, improving user experience for large models on slow storage.

Health Assessment

• • Fast turnaround with a single commit and review, minimal rework, and a focused fix that improves stability and performance for macOS users loading large models.

Summary

Add structured output support to MLX runner

The MLX runner now enforces JSON and JSON Schema formats, preventing unconstrained text outputs. This enhances API reliability for structured data generation, aligning with user expectations and reducing errors.

Health Assessment

• • The PR introduced significant new functionality with moderate review time and a few iterations, indicating a well‑managed process. The large code additions suggest careful testing and documentation to support the new structured output feature.

Summary

Update llama.cpp version to b10630

Bumps the llama.cpp dependency to a newer commit, ensuring compatibility and bug fixes.

Health Assessment

• • Quick review and merge, minimal changes, low risk.

Summary

MLX version bump

Update the MLX dependency to the latest release, ensuring compatibility and security.

Health Assessment

• • Version bump with minimal changes; review took over a day but overall low risk and impact.

Summary

Make external compat patches idempotent

Ensures MLX-C rebuilds succeed by applying patches idempotently, preventing build failures on macOS.

Health Assessment

• • Single AI-assisted commit, moderate review delay, small code change, low risk to production

Summary

Preserve string content during image fallback

Ensures Claude Desktop handles string-form content correctly, preventing 400 errors when unsupported images are encountered.

Health Assessment

• • Quick fix with minimal changes, resolved in under 15 minutes with no review iterations.

Summary

Reset Claude Desktop models to defaults

Users can now reset Claude Desktop models to the correct defaults, improving consistency and reducing configuration errors. Windows users no longer see irrelevant settings.

Health Assessment

• • PR merged quickly with a single review, indicating straightforward changes, but the large code churn suggests a high-impact update.

Summary

Keep Claude toggle active while connecting

Ensures users see a consistent UI state during the Claude enable flow, reducing confusion and improving perceived responsiveness.

Health Assessment

• • Fast review and merge within 36 minutes, minimal changes, low risk.

Summary

Reject browser origins on Claude Desktop gateway

Prevent browsers from accessing the Claude Desktop gateway by rejecting origins, ensuring OLLAMA_CORS does not affect port 11435.

Health Assessment

• • Quick review and merge indicate a straightforward, low‑risk change with minimal impact on the codebase.

Summary

Health Assessment

Summary

Health Assessment

Summary

Health Assessment

Summary

Health Assessment

Summary

Register DeepSeek Harness in docs

Adds the DeepSeek Harness guide to the documentation sidebar, enabling users to access it via the Integrations navigation.

Health Assessment

• • Quick documentation update with minimal changes, merged within 4 hours, indicating efficient review and low complexity.

Summary

Tolerate non-leading system messages in Qwen renderer

Allows Qwen renderer to handle runtime system messages after the initial user turn, preventing HTTP 500 errors and improving robustness for coding clients.

Health Assessment

• • Rapid approval and merge with minimal changes indicates low complexity and risk.

Summary

Transcode WebP images for llama-server

Adds WebP support by detecting and converting WebP images to PNG before forwarding, expanding image format coverage for llama-server.

Health Assessment

• • Rapid review and merge indicate a straightforward change with minimal friction.

Summary

Add developer instruction support for Qwen3.8

Enables Qwen3.8 to handle developer instructions and expands streaming tool‑call coverage across multiple request shapes, improving compatibility with OpenAI and Anthropic APIs.

Health Assessment

• • Rapid review and merge within 1.1 hours indicates minimal friction and high confidence in the change.

Summary

Add Qwen3.8 renderer and MLX import support

Adds support for Qwen3.8 model, enabling advanced chat template semantics. Improves safetensors import handling and quantization support, enhancing compatibility and performance.

Health Assessment

• • Fast review and merge within 45 minutes, indicating high confidence and minimal friction.

Summary

Add DeepSeek Harness integration

Enables users to launch DeepSeek Harness models directly from Ollama, expanding first‑party model support and improving workflow.

Health Assessment

• • Fast turnaround with minimal rework; PR merged within 3 hours, indicating high confidence and low complexity.

Summary

Update Muse Glimmer renderer and template

Ensures consistent reasoning prompts by normalizing and deduplicating system directives, improving user experience and preventing duplicate reasoning lines.

Health Assessment

• • Fast 2.5‑hour cycle with no rework indicates a well‑scoped, low‑risk change.

Summary

Add Muse Code integration to launch

Adds support for Meta's Muse Code CLI, allowing users to launch Muse Code directly from Ollama. This streamlines model deployment and integration for developers.

Health Assessment

• • The PR had a quick initial review but a long overall cycle time due to a delayed final commit. The large code addition and extended timeline increase risk for integration issues.