Intel Arc Pro B70 for local LLMs: measured results

8 results from 3 builds (one and two cards), each linked to its source.

All measured B70 results

Hide speculative-decoding results ⚡ (MTP, DFlash, DSpark and similar draft-and-verify methods give much higher tok/s on predictable text and are not comparable 1:1 with plain decode)

⏳ Pending: neural.download marks this headline 'pending': speed measured, install checks not finished.

What fits in 32 GB and 64 GB

Q4_K_M weights (llama.cpp quantize table), before KV cache. One card (32 GB): Flint 40B-A2.5B ≈24 GB, Ornith 1.5 35B-A3B ≈22 GB, Qwen3.6 35B-A3B ≈22 GB, Gemma 4 31B ≈19 GB, Nemotron 3.5 Lightning 30B-A3B ≈18 GB, Qwen3.6 27B ≈17 GB, Qwen3.8 27B ≈17 GB, Gemma 4 26B-A4B ≈15 GB, Devstral Small 2 24B ≈15 GB, LFM2.5 24B-A2B ≈15 GB. Two cards (64 GB): Kolibri-1 ≈48 GB, Nemotron 3.5 Bolt 74B-A9B ≈45 GB, Llama 3.3 70B ≈43 GB.

Price

We do not track a B70 street price yet, so there is no tok/s-per-euro figure here. Compare with the Radeon AI Pro R9700 (also 32 GB) and the RTX 5090.

More: all Arc Pro B70 builds · best LLMs for 32 GB · 64 GB

FAQ

Is the Intel Arc Pro B70 good for local LLMs?

It has 32 GB per card. RigBase has 8 measured results from 3 B70 builds; the fastest single-card plain-decode result is 122.16 tok/s on Gemma 4 26B-A4B (UD-Q8_K_XL). 3 of the 8 results use speculative decoding, so compare plain with plain.

Which software runs LLMs on the Arc Pro B70?

In the builds we track: llama.cpp SYCL, neural.download runtime, vLLM XPU. Speeds differ a lot between stacks for the same model, so check the engine column.

What fits in one or two B70s?

By Q4_K_M memory math (weights only, before KV cache), one 32 GB card fits models up to about 30 GB of weights, e.g. Flint 40B-A2.5B, Ornith 1.5 35B-A3B, Qwen3.6 35B-A3B, Gemma 4 31B; two cards (64 GB) add Kolibri-1, Nemotron 3.5 Bolt 74B-A9B, Llama 3.3 70B.

Why is one result marked pending?

The source (neural.download) marks its Ornith 1.5 35B-A3B headline as pending: speed measured, install checks not finished. We show it with that label and leave it out of “fastest” summaries.