Announcements

Cline 3.38.0: Gemini 3 Pro Preview & Voice Dictations that understands your code

Cline 3.38.0 introduces Gemini 3 Pro Preview for enhanced reasoning and multimodal capabilities. We've also added AquaVoice Avalon, a voice-to-text model that specifically understands technical terms

Juan Pablo Flores

17 Nov 2025 — 2 min read

Cline 3.38.0 introduces the Gemini 3 Pro Preview model for enhanced reasoning and multimodal capabilities, alongside the Aquavoice Avalon model for voice-to-text dictation that specifically understands technical terms. This release also includes critical fixes for AWS Bedrock context handling and tool usage optimizations to prevent premature context truncation.

Gemini 3 Pro Preview

Gemini 3 Pro is Google’s new flagship model for reasoning, multimodal work, and coding, is available now in Cline. In our testing, it represents a clear step up from Gemini 2.5 Pro, showing top scores on coding tasks, agentic workflows and reasoning benchmarks.

Within Cline, this translates to a model that holds onto structure over longer tasks, sticking to the original plan instead of wandering into unrelated edits. It makes genuine use of its 1M-token context window to respect existing patterns and keep changes coherent across multiple files. We've also seen impressive frontend generation capabilities that rival other SOTA models, especially when mixing code, design screenshots, and documentation in one conversation.

0:00

/0:43

We’ve written a full deep-dive on what this means for your workflow, which you can read here: Gemini 3 Pro gives Cline more headroom for long-running coding.

Voice Coding with AquaVoice Avalon

Our team is often talking to Cline, but generic speech models often struggle with the specific language of software development. We’ve integrated Avalon from AquaVoice, a speech model tuned specifically for how engineers actually talk.

Instead of misinterpreting commands like "checkout dev", Avalon gets it right. On the AISpeak benchmark of coding terms, Avalon achieves 97.4% accuracy compared to Whisper Large v3's 65.1%. If you use voice-first coding, you’ll notice the difference immediately.

0:00

/2:10

Fixes

Automatic context truncation for AWS Bedrock: Fixed an issue where context truncation wasn't triggering correctly when token usage rate limits were exceeded.
Optimized new_task tool usage: We removed the new_task tool from system prompts and updated slash command prompts. We noticed that some models, particularly Sonnet 4.5, were "context anxious" and triggering auto-condensing prematurely. This change prevents unhelpful early usage while keeping the tool available via slash commands (/new_task) or custom rules.
Native tool calling validation: Added a helper function to ensure more robust native tool calling.