#2364: Running Hermes Agent on ARM64 mobile via AVF — real-world performance data point

Title: Running Hermes Agent on ARM64 mobile via AVF — real-world performance data point

Description:

Hi Hermes team,

I've been running Hermes Agent on an unconventional setup — a mid-range Android phone via AVF (Android Virtualization Framework) — and wanted to share real-world performance data that might be useful for ARM64 optimization and mobile use cases.

Hardware used:

• Device: Xiaomi Redmi Note 13 Pro 4G (model 23117RA68G)
• SoC: MediaTek Helio G99-Ultra (6nm, 8-core: 2x Cortex-A76 @ 2.2GHz + 6x Cortex-A55 @ 2.0GHz)
• RAM: 8 GB physical (no memory extension active)
• Storage: 256 GB
• OS: Xiaomi HyperOS 3.0.2.0 (Android 16)

Execution environment:

• Hermes runs inside a Debian 12 (Bookworm) VM via AVF
• VM is allocated 3.8 GiB RAM (out of 8 GB physical)
• 8 CPU cores visible to the VM
• Kernel: 6.1.0-34-avf-arm64
• Disk: 46 GB virtual (29 GB free)
• Memory at rest: ~694 MB used (82% available)
• Swap: 981 MiB (unused)

Hermes versions:

• Agent CLI: v0.13.0 (2026.5.7)
• WebUI: stage-362 (commit 352064e)

Setup details:

• All inference is cloud-based (NVIDIA NIM free tier + OpenCode Zen free tier)
• Accessed via Chrome on Android at localhost:8787 (standard tab, not PWA)
• Performance mode: OFF
• Battery optimization: OFF for Hermes (to prevent background killing)
• Custom persistence script prevents Android from terminating terminal threads
• Connection: 4G Plus mobile data

Performance findings:

Provider	Peak speed	Notes
OpenCode Zen (DeepSeek Free)	819 t/s 🏆	Peak observed under normal 4G+ conditions
Nous (DeepSeek V4 Flash)	—	No tok/s metric exposed by Nous provider

Benchmark: Nous DeepSeek V4 Flash vs OpenCode Zen DeepSeek Free (May 16, 2026):

Test	Nous (v4 Flash)	OpenCode Zen (Free)	Difference
Letter counting (reasoning)	51.0s	35.4s	31% faster
Fibonacci DP (code generation)	45.8s	38.1s	17% faster
Average	48.4s	36.8s	~24% faster

Both models produced correct results. OpenCode Zen Free was consistently faster.

Thermal behavior (Helio G99-Ultra):

Condition	Temperature
Normal Hermes usage	✅ Cool — no heat
Charging + Wi-Fi hotspot active	🔥 Warm (external load, not AI-related)
Charging alone or hotspot alone	✅ Normal

The SoC handles AI inference streaming without overheating. Heat only appears under combined external loads (charging + hotspot).

Context length degradation:

• Short/medium conversations: Stable — no speed loss
• Extremely long contexts: ~30% estimated degradation

Hermes footprint:

• Agent install: 1.3 GB
• WebUI: 220 MB
• Config/sessions/skills/logs: 199 MB
• Total: ~1.7 GB

What works well:

• ARM64 architecture is fully compatible
• AVF provides stable virtualization for the Debian environment
• 3.8 GiB is sufficient RAM for Hermes + background processes
• Chrome renders the WebUI smoothly even at +800 t/s streaming
• The mid-range Helio G99-Ultra handles inference rendering without lag

Known limitations / observations:

1. Model cache confusion — models_cache.json has a 24h TTL and persists on disk across restarts. Adding new models to _PROVIDER_MODELS requires manually deleting the cache file (rm -f ~/.hermes/webui/models_cache.json) — the UI won't pick up changes otherwise until the TTL expires.
2. PWA / native app gap — on mobile, switching to another app (gallery, WhatsApp, browser) causes the Chrome tab to reload, interrupting the session. The session recovers (Hermes persists state), but it disrupts workflow. A robust PWA or native wrapper would solve this.
3. No ARM64 pre-built wheels — some dependencies compile from source on first install, which adds setup time on the phone.

Questions for the team:

1. Are there any known plans for ARM64-specific optimizations or pre-built wheels?
2. Has anyone else tested Hermes via AVF or similar mobile virtualization?
3. Would a PWA manifest be accepted upstream to improve mobile UX?

Happy to provide more data, run specific tests, or help with ARM64 compatibility testing. Running Hermes on a phone feels like the future — this thing works surprisingly well on mid-range mobile hardware.

Keyos/