Question 1

What models can the Intel Arc 140V (Lunar Lake) run?

Accepted Answer

With ~13.5GB of usable memory: qwen3:0.6b, qwen3:1.7b, qwen3:4b, qwen3:8b, qwen3:14b, qwen2.5vl:7b, llama3.1:8b, llama3.2:1b run fully in memory. Soldered LPDDR5X-8533 on package — RAM is NEVER upgradeable; 16 GB 1-channel-single-rank SKUs drop to ~68 GB/s vs 136 GB/s on 32 GB dual-rank SKUs (wccftech SKU table), so check rank before quoting bandwidth llama.cpp works via SYCL (IPEX-LLM / oneAPI) or Vulkan, not CUDA — needs Intel oneAPI runtime on Linux; community bench: Llama 2 7B Q4_0 SYCL ~14.4 tok/s tg128, ~180 tok/s pp512 (llm-tracker.info Xe2 writeup) SYCL F16 backend crashed on k-quants in some IPEX-LLM builds (intel/ipex-llm issue #12318) — stick to Q4_0 on early stacks Arc 130V (Core Ultra 5 SKUs) is the same chip with 7 Xe cores at 1.85 GHz — slightly slower, same memory platform XMX 'tensor cores' (128 XMX engines) accelerate prefill via SYCL; decode tok/s is memory-bound like any iGPU

Question 2

How fast is the Intel Arc 140V (Lunar Lake) for local LLMs?

Accepted Answer

computed band 6.6-21.0 tok/s for 8B Q4_K_M (roofline, 136.5 GB/s shared RAM); community: tg ~14.4 / pp ~181 tok/s Llama 2 7B Q4_0 (https://llm-tracker.info/howto/Intel-GPUs)

Question 3

Is the Intel Arc 140V (Lunar Lake) still worth using for local LLMs in 2026?

Accepted Answer

computed band 6.6-21.0 tok/s for 8B Q4_K_M (roofline, 136.5 GB/s shared RAM); community: tg ~14.4 / pp ~181 tok/s Llama 2 7B Q4_0 (https://llm-tracker.info/howto/Intel-GPUs) Soldered LPDDR5X-8533 on package — RAM is NEVER upgradeable; 16 GB 1-channel-single-rank SKUs drop to ~68 GB/s vs 136 GB/s on 32 GB dual-rank SKUs (wccftech SKU table), so check rank before quoting bandwidth llama.cpp works via SYCL (IPEX-LLM / oneAPI) or Vulkan, not CUDA — needs Intel oneAPI runtime on Linux; community bench: Llama 2 7B Q4_0 SYCL ~14.4 tok/s tg128, ~180 tok/s pp512 (llm-tracker.info Xe2 writeup) SYCL F16 backend crashed on k-quants in some IPEX-LLM builds (intel/ipex-llm issue #12318) — stick to Q4_0 on early stacks Arc 130V (Core Ultra 5 SKUs) is the same chip with 7 Xe cores at 1.85 GHz — slightly slower, same memory platform XMX 'tensor cores' (128 XMX engines) accelerate prefill via SYCL; decode tok/s is memory-bound like any iGPU For serious work, any modern entry-level dGPU will do.

Best local LLMs for the Intel Arc 140V (Lunar Lake)

Speed

What fits (computed)

Cited community benches (1 rows)

Runs fully in memory (machine alternatives)

Borderline — runs, but offloads

Gotchas

Where to go next