Viper releases on-device AI chat feature on Mac
Understand this story
Viper has launched an AI chat feature that runs entirely locally on the Mac, based on gpt-oss-20b; it does not modify the model weights, but instead injects application knowledge and constrains responses through an external harness. It currently supports only Apple silicon Macs with 24 GB of memory; Viper says the iPhone is not yet supported because its maximum memory of 12 GB is below what the model itself requires. Viper says that on 167 new blind-test questions, version 1.5.35 answered correctly 61% of the time and gave wrong or fabricated content only 4% of the time, while the bare model answered correctly 2% of the time with a fabrication rate of 57%. Viper notes that these figures come from the test bench, not from the app build. Version 1.5.34 has entered App Store review; all actions are presented as drafts and require user confirmation before being sent.
AI overview based on source reports · Updated 2 days ago
Coverage timeline
Follow the coverage to explore different perspectives.
- Hacker News AIOn-device AI: new knowledge without fine-tuning
Viper describes its Mac AI chat that runs an unchanged open model (gpt-oss-20b) entirely on-device, teaching it about the app through a harness of guides, question splitting and code checks rather than fine-tuning. On 167 fresh blind-judged questions, the harness improved correct answers from 2% (model alone) to 41% (1.5.34) and 61% (1.5.35), while invented answers fell from 57% to 24% and 4%. It requires an Apple silicon Mac with about 24 GB of memory, and the 1.5.35 figures come from the test bench, not yet an app build.
Activity for this story
- Current comparable activity
- 4
- Peak comparable activity
- 10Nov 19 08:00
- Change in the past 24 hours
- -50%
Trends compare the same fully observed participants, which may cover fewer participants than current activity. Move the pointer or tap the chart to inspect hourly activity; use the left and right arrow keys with a keyboard.