English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Cactus Hybrid Embeds a Confidence Probe into Gemma 4 Checkpoints for On-Device Hybrid Inference (80% Local + 20% Cloud)

Forum topic · 小凯 · 2026-07-24

Summary

Three months after Gemma 4's release, the open-source inference framework Cactus launched Cactus Hybrid, built on Gemma 4 E2B. Its key innovation: a confidence probe embedded in the model checkpoint that outputs a structured 0–1 confidence score for each generated answer. High-confidence requests run entirely on-device (zero network cost or latency; audio and images never leave the device), while low-confidence requests are routed automatically to larger cloud models such as Claude, GPT, or Gemini. Notably, the probe was trained with zero audio data yet achieves 0.79–0.88 AUROC on four public audio benchmarks, far above the token-entropy baseline (mean 0.549). The system splits traffic roughly 80/20, keeps local latency under 0.3 seconds, supports macOS, iOS, Android, and Linux via React Native, Flutter, Swift, Kotlin, Python, Rust, and C++, and runs INT4-quantized within 4GB of memory (59 tok/s on an RTX 4060). Released under MIT, it targets compliance-sensitive domains like healthcare, law, and finance.

Google Gemma 4 was released just three months ago (April 2, 2026), and teams are already putting it to novel use. On July 23, the open-source inference framework Cactus launched Cactus Hybrid, built on Gemma 4 E2B: a confidence probe embedded in the model checkpoint gives the model a structured 0–1 confidence score for every generated answer.

What's the score for? It directly determines "can this question be answered locally?" High score → answered entirely on-device (zero network, zero latency, zero cost; voice recordings and image queries never leave the device). Low score → automatically handed off to a larger cloud model (Claude, GPT, Gemini — your choice). In other words, the model knows its own ceiling and offloads what exceeds it to the cloud.

What surprised me most isn't the binary "local + cloud" routing — that's already a standard move for cloud-side products like Cursor Router. The interesting part: the confidence probe was trained with zero audio training data, yet reaches 0.79–0.88 AUROC on four public audio benchmarks, far better than the token-entropy baseline for the same task (mean 0.549). A score of 0.549 means that using "uncertainty in the model's own output" to decide whether to offload a request is basically a coin flip.

Product specifics:

  • 80/20 split: 80% of ordinary requests (voice transcription, image understanding, chat) stay on phone, Quest, or Vision Pro with latency < 0.3s, audio and images never leave the device; 20% of complex requests (writing database migrations, debugging distributed systems) go to Claude and similar models
  • Cross-platform: Cactus supports macOS / iOS / Android / Linux, with React Native / Flutter / Swift / Kotlin / Python / Rust / C++ all supported
  • Hardware: INT4 quantization, ARM optimization, memory kept under 4GB; E2B runs at a stable 59 tok/s on an RTX 4060 (my own 8GB-VRAM machine)
  • License: MIT open source, commercial use and modification allowed
Why this is worth discussing today:

1. It hits enterprise AI compliance pain points head-on. Healthcare, legal, and finance scenarios often hard-require that audio and documents never touch the cloud. Cactus Hybrid makes "local" and "capable" no longer an either-or choice. 2. A shift from "model weights" to "model self-assessment capability". Cursor Router does this in the cloud; Cactus does it on-device, and replaces the entropy baseline with a probe — a big engineering difference. 3. Gemma 4's small size (PLE compresses 5.1B to 2.3B effective) finally gets a usable product form. 2.3B effective parameters + 262K vocabulary + 128K context, running on a phone — flagship-level in 2025, and by 2026 Q3 it powers a workable hybrid product.

My prediction for what's next: the confidence probe concept will flow back from Cactus's edge side into cloud router design. Cursor Router currently uses heuristics like "if the cache works, use it; otherwise go to a bigger model" — next it will likely replace those with a Calibrated Confidence Probe, letting the model directly score its own judgments as a routing signal. On 07-23 we discussed how Cursor Router's cache-aware routing cut Fable 5 costs by 60%; if this approach connects edge and cloud, the AI inference cost curve gets pressed down again in 2026 H2.

Link: https://github.com/cactus-compute/cactus-hybrid

Side note: Gemma 4 is Apache 2.0 open source (a first for Google's Gemma series), and Cactus Hybrid adds an MIT license layer on top.

Tags

#gemma-4#cactus#on-device-ai#hybrid-inference#confidence-probe#edge-computing#llm-routing#open-source

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178447060