Holo3 · OSWorld · 2×A6000

Progress

–

tasks done / total

–

success rate

–

scored

–

elapsed

–

ETA remaining

–

mean task wall

official Holo3-35B-A3B leaderboard runs: 82.56 / 78.15 % (mean 80.4) at the same 100-step budget · sibling run: NUC16 (CPU edge)

Current task

idle

Live system telemetry

Serving aggregate

Success rate by domain

GPU power (W) & GPU util — last ~3 h

Tasks

Engine bring-up — llama.cpp CUDA vs vLLM (evidence)

…

The four regimes, inverted — edge (NUC16) vs datacenter (2×A6000), live

Same model, same screenshot-only agent, same OSWorld tasks — read against the KV-cache four-regimes framework. On the edge the agent step is prefill-bound (re-prefilling the evicted context dominates), which is the whole reason the evict / no-evict / KV-optimum tradeoff exists. On the A6000 prefill collapses to ~1–2 s, the step becomes VM/environment-bound, and the KV-cache regime is moot (vLLM also has no prefix cache for this hybrid arch). The bottleneck — and the energy story — invert. Numbers update live from both runs' state feeds.

Same model, three devices — what changes and why

This run (2×RTX A6000, GDDR6 ~768 GB/s each) — the same hybrid qwen3_5_moe (Gated-DeltaNet × 256-expert MoE VLM) that limps on edge silicon simply works on discrete CUDA:

The Vulkan failure classes don't reproduce on CUDA. Full offload (-ngl 99) decodes correct tokens at temp 0 (the Xe3 lm_head corruption is a Vulkan kernel bug, not a model property), and the CLIP/mmproj vision encoder runs clean (no DeviceLost) — UI localization lands within ~2 px.
Bandwidth is the whole story for MoE decode. 3B active parameters streamed from 768 GB/s GDDR6 instead of ~90 GB/s shared DDR5 → decode ~10× the NUC (and prefill ~28–75×, compute-bound).
What stays identical: task set (361 = 369 − 8 Google-Drive), max_steps=100, screenshot-only observation, temp 0.8, 1920×1080, single rollout, H Company agent-loop protocol.

Why this comparison matters. NUC16 measures what a fanless edge box can do today; this run measures the same agent with the serving bottleneck removed — isolating model capability from serving physics. Score deltas between the two runs bound the quality cost of edge serving (quant + latency-induced behavior differences), while the step-latency/energy gap quantifies its price.