Builds / Intel Arc 140V iGPU (Core Ultra 7 258V, Lunar Lake, shared m
ThinkPad X1 Carbon Gen 13 (Core Ultra 7 258V, Arc 140V iGPU, 32 GB): Qwen3.5 35B-A3B IQ4_XS at 25.6 tok/s on llama.cpp Vulkan, 22.3 at 8K depth
Last updated:
- GPU(s)
- Intel Arc 140V iGPU (Core Ultra 7 258V, Lunar Lake, shared memory)
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 32
- CPU
- Intel Core Ultra 7 258V (Lunar Lake)
Amazon · eBay (used) · Newegg - Motherboard
- Lenovo ThinkPad X1 Carbon Gen 13 (laptop)
Amazon · eBay (used) · Newegg - RAM
- 32 GB LPDDR5X-8533 (shared CPU/GPU)
Amazon · eBay (used) · Newegg - PSU
- not stated
- Other
- Arch Linux, Mesa 26.2.7 (Intel open-source Vulkan, KHR_coopmat); llama.cpp build 32bdc3a89 (12377); powerprofilesctl balanced; MoE runs used 2 repetitions to limit thermal throttling
- Reported total cost
- not stated
- Power
- laptop, balanced power profile
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Qwen3.5 35B-A3B legacy | IQ4_XS (16.95 GiB) | llama.cpp 12377 Vulkan (Mesa ANV) | 25.58 | 379 @ 8192-token prompt est. TTFT @8k ≈ 21.6 s (calculated from measured prefill) | context: llama-bench tg128📏 -ngl 99 -t 4 -fa 1; tg128 25.58, 22.31 ± 0.37 at 8,192-token depth; pp8192 379.11 with -b 4096 -ub 1024 |
| Llama 2 7B legacy | Q4_0 | llama.cpp 12377 Vulkan (Mesa ANV) | 24.25 | 667 @ 512-token prompt est. TTFT @8k ≈ 12.3 s (calculated from measured prefill) | context: llama-bench tg128📏 FA off: tg128 24.25 ± 0.03, pp512 667.34; FA on was slower (18.70 / 619.20) |
Prompt processing = prefill speed (how fast the prompt is read) as reported by the source, with the prompt length it was measured at; blank = the source does not report it. est. TTFT @8k is calculated from the measured prefill speed (8,192 ÷ prefill tok/s), shown only when prefill was measured on a prompt of 512+ tokens. It is not a measurement and ignores load time and caching. measured TTFT is the source's own time-to-first-token. Why prefill matters for coding assistants.
Owner's run in llama.cpp's Vulkan scoreboard discussion (14 Nov 2026). A thin-and-light laptop with 32 GB of shared memory runs a 35B MoE (3B active) at about the same speed as a dense 7B, because only the active experts are read per token. The 17 GB model leaves little room on a 32 GB machine, and laptop power/thermal limits apply.
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.