Hermes Agent Local Model Benchmark: ~93 tok/s on RTX 5090

Real one-hour Hermes Agent goal run on Qwen3.8-27B (6-bit, MTP draft, q8_0 cache, 128k context) averaging 93 tok/s generated (297k tokens). Prompt processing ~2.1–2.2k tok/s. Fits comfortably on 32 GB; 192k also possible. Output was a full voxel diorama of Mont-Saint-Michel in one HTML file.

Credit@witcheer

Cover art for Community: Hermes Agent Local Model Benchmark: ~93 tok/s on RTX 5090

Why it matters

Concrete evidence that serious local agent work is already viable on consumer hardware, with exact setup notes others can replicate.

From the source

Post on X@witcheer

Hermes Agent #local-models #llama.cpp #rtx-5090 #performance #hermes-agent #qwen

More Community

CommunityNov 1

Multiplex vs Separate Gateways for Multi-Profile Hermes

Detailed trade-off analysis of running one multiplexed gateway (gateway.multiplex_profiles true) versus a separate gateway per agent profile. Covers failure domains, cron duplication, blast radius of runaway loops, log isolation, and the real meaning of isolation (process count is not a security boundary). Recommendation: multiplex unless a profile has its own channel, needs pinned runtime, separate user/limits, or must survive others’ restarts.

Credit@O_Piekielniak

CommunityOct 31

ClawPanel: Multi-Engine Control Panel

Open-source multi-engine panel supporting OpenClaw, Hermes Agent, DeepSeek Harness, and OpenCode. Aimed at turning agent demos into usable systems with unified management. ~3k stars.

Credit@JayQinQT