#moe

LFM2.5-8B-A1B (Q4_K_M, 8.47B total / ~1.7B active MoE params, 4.79 GiB) measured on TWO GPUs for a clean MoE-vs-GPU comparison: EvoX2's AMD Radeon RX 7900 XT (gfx1100, ROCm) at 269.01 tok/s gen / 7464.85 tok/s prompt using 6.81GB VRAM, and Jitori's Intel Arc Pro B60 (Battlemage, Vulkan) at 68.79 tok/s gen / 1688.53 tok/s prompt. The RX 7900 XT is ~3.9x faster on this MoE than the Arc B60 - the SAME ratio as on the dense Llama-3.1-8B (see bm-008), confirming the GPU gap is consistent across architectures. Companion to bm-011 (the APU iGPU result on the same machine). Closes the small-model and real-app workload gaps.

LFM2.5-8B-A1B (Q4_K_M, 8.47B params, 4.79 GiB) on the integrated Radeon 8060S GPU of EvoX2's AMD Ryzen AI MAX+ 395 (gfx1151, ~124GB unified memory pool visible to HIP). Generates at 150.16 tok/s and processes prompts at 3661.53 tok/s via ROCm 7.2.2. This is the only unified-memory APU in the fleet - the same machine's discrete RX 7900 XT (bm-007) is ~1.8x faster on this small model, but the APU's point is not raw speed: unified memory removes the VRAM ceiling that bounds every discrete card here. Closes the unified-memory / APU platform gap and answers the 'should I buy Strix Halo / Ryzen AI MAX for local AI?' question with first-party data.

Qwen3-30B-A3B-Instruct-2507 (INT4, OpenVINO IR) served via OpenVINO Model Server 2026.2.1 on Jitori's Intel Arc Pro B60 (Battlemage BMG G21, 24GB). Generates at 67.95 tok/s via the OpenAI streaming API - roughly 1.76x faster than the same B60 running Qwen3-Coder-30B-A3B Q4_K_M via llama.cpp Vulkan (bm-001, 38.6 tok/s). This is the backend comparison the dataset was missing: same GPU, same 30B-A3B model class, two Intel-GPU backends. Finding: OpenVINO is meaningfully faster than Vulkan on Battlemage for this workload. This also corrects a fleet assumption - OpenVINO's GPU plugin (Level Zero) works cleanly on this B60, contradicting the older 'SYCL is broken on Battlemage, Vulkan-only' note.

For the Ornith-1.0-35B MoE model (qwen35moe family) on a single Radeon AI PRO R9700, Ollama is the clear winner at 78.3 tok/s - roughly 2.9x faster than llama.cpp Vulkan (27.3 tok/s), which also spilled past the 32GB VRAM limit into system RAM. The HIP backend failed outright: it could not parse the Ornith GGUF (version mismatch). On MoE models without an MTP head, Ollama's expert routing currently beats hand-tuned llama.cpp on this build.