#benchmarks

A year ago Intel Arc was still a punchline for AI workloads. The Battlemage refresh (B580/B60) is the thing that changed that — and a year of driver, backend, and ecosystem work has made it genuinely good, with caveats. Here's where Arc actually stands for local AI in mid-2026.

Every 'best local LLM' list ranks models by reputation and parameter count. None of them actually ran the models on consumer hardware and reported measured tok/s, VRAM, and quality scores. We did. Here are the models that genuinely run well on hardware you own — with the receipts.

In 2024, Q4_K_M meant 'noticeably broken.' In 2026 it's the practical default. We ran the Q4 vs Q6 vs Q8 sweep on Qwen3-Coder-30B — here's the measured quality delta, the VRAM math, and when you actually should step up.

Every model worth running locally in 2026 — Qwen3-30B-A3B, LFM2.5-8B-A1B, the GLM/Kimi frontier — is MoE. Here's what 'active parameters' actually means, why it's the reason these models fit on your card, and the measured 2.5× speed data behind it.

Every GPU roundup ranks the same NVIDIA cards. Almost none run the same model file on Arc, AMD, and NVIDIA and report the actual tok/s gap. We did. The cross-vendor gap is real. So is the price gap — and that flips the recommendation for budget buyers.

Every coding-LLM roundup ranks models by leaderboard score. Almost none tell you what the score costs in tok/s when you step down a quant tier — which is the only trade-off that matters once the model has to fit on a card you own. Here's the measured table for the coder we run, across three quant tiers, on the same hardware.

We took Qwen3.6-27B at Q4_K_M — byte-identical weights, verified by Ollama digest — and ran it on all four lab machines: dual 12GB NVIDIA, single 20GB AMD, dual 32GB AMD, and a 24GB Intel Arc. The three CUDA/ROCm machines landed within 10% of each other. The honest answer to 'do I need two GPUs?' is: only once the model stops fitting in one.

Every backend explainer describes CUDA, ROCm, Vulkan, and SYCL correctly and then tells you to 'pick what works on your card.' Almost none measure them on the same hardware. We did, across nine benchmark records. The backend gaps are large, vendor lock-in is real, and one documented failure is worth a thousand abstract paragraphs.

The 'Ollama is easier, llama.cpp is faster' framing is from 2024 and it was correct then. What it never came with was the actual measured gap on the same card with the same weights. Here's that gap — across four backend configurations on a single R9700, plus the one result that looks like an Ollama win and isn't.

Every 'how much power does local AI use' piece multiplies GPU TDP specs by hours and calls it a day. Almost nobody measures actual wall-power during inference. We measured a narrow slice — and the headline finding is that quantization barely moves power draw. The honest version of what we have, what we don't, and the rough cost math.

Forum threads answer 'is dual GPU worth it' with 'depends.' Our four-machine baseline (bm-013) gives the actual crossover: dual-12GB-NVIDIA tied single-32GB-AMD on a 27B model. Below ~18GB of model weight the second card idles; above, dual becomes mandatory or faster. The measured rule, with the trade-offs single-card buyers forget.

Every speculative decoding explainer calls it a 'free speedup.' Almost none measure the actual multiplier on consumer hardware or show the accept-rate curve across context lengths. Here's both — Qwen3.6-27B on the same R9700 with and without MTP, plus the long-context sweep that explains when the speedup holds and when it shrinks.

We ran the same two 8B models on both cards with byte-identical weights and identical settings. The RX 7900 XT is ~3.9x faster — consistently across both architectures. Here's the full 2x2, what it actually tells you, and where we got the framing wrong the first time.

We benchmarked OpenVINO Model Server 2026.2.1 against llama.cpp Vulkan on the same Intel Arc Pro B60. OpenVINO is ~1.76x faster on a 30B-A3B model — and it works cleanly, contradicting our own earlier note that Battlemage was Vulkan-only. Here's the data and the correction.

Yes, really. A Q4_K_M 30B coder model fits in 12GB of VRAM and runs at usable speed on a single Intel Arc B580. Here's the recipe — with the numbers to back it.