Builds / RTX 5060 Ti 16GB (eGPU, OcuLink PCIe Gen4 x4) + Radeon 780M
Mini PC (Ryzen 7 255, 64 GB) + RTX 5060 Ti 16 GB over OcuLink: Qwen3.8 Flash-Next Q2_0 on Strata at 93 tok/s, Qwen3.8 27B at 64K with MTP 45.7 (Somenzi)
Last updated:
- GPU(s)
- RTX 5060 Ti 16GB (eGPU, OcuLink PCIe Gen4 x4) + Radeon 780M iGPU
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 16
- CPU
- AMD Ryzen 7 255
Amazon · eBay (used) · Newegg - Motherboard
- Mini PC, M.2-to-OcuLink adapter
Amazon · eBay (used) · Newegg - RAM
- 64 GB DDR5
Amazon · eBay (used) · Newegg - PSU
- AOOSTAR AG02 dock (800 W) for the GPU
Amazon · eBay (used) · Newegg - Other
- Omarchy Linux, nvidia-open-dkms; desktop stays on the Radeon 780M, the RTX 5060 Ti only computes; link measured 7.2 GB/s each way
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Qwen3.8 Flash-Next | Strata Q2_0 (66 GB on disk, 34 GB experts), experts split VRAM/RAM, KV in RAM | Strata engine 0.1.54 (prebuilt CUDA) | 93 | 1,613 @ 16K-token prompt est. TTFT @8k ≈ 5.1 s (calculated from measured prefill) | ⚡ spec-decode (not stated (Strata defaults))context: generation at 16K of context, 128K configured📏 Somenzi run 4 Nov 2026, published 10 Nov: three runs with a fresh server each, 93 to 97 tok/s (lowest listed); long code test 108 to 112 tok/s; an opencode agent task averaged 87 tok/s over 135 requests |
| Qwen3.8 27B | IQ3_S + MTP head (11.3 GiB), q8_0 KV | llama.cpp (CUDA, built 11 Nov 2026) | 22.7 | 824 @ 56K-token prompt est. TTFT @8k ≈ 9.9 s (calculated from measured prefill) | context: 65K ctx, 56,066-token agent-like context📏 Somenzi llama-server bench (12 Nov 2026): fresh server per variant, 3 rounds, temperature 0, within 0.5%; same on new code, edit and plain text |
| Qwen3.8 27B | IQ3_S + MTP head (11.3 GiB), q8_0 KV | llama.cpp (CUDA, built 11 Nov 2026) | 45.7 | not stated | ⚡ spec-decode (MTP)context: 65K ctx, 56,066-token context📏 Writing new code; 51.5 on an edit task, 32.6 on plain text (666 of 1,305 drafts accepted); output code identical to no speculation; cold prefill drops to 777 tok/s with MTP on |
| Qwen3.8 27B | IQ3_S + MTP head (11.3 GiB), q8_0 KV | llama.cpp (CUDA, built 11 Nov 2026) | 131.7 | not stated | ⚡ spec-decode (MTP + n-gram)context: 65K ctx, edit task📏 --spec-type draft-mtp,ngram-mod on an edit (returning an existing function with one variable renamed), where the answer mostly copies the context; new code 45.7 and plain text 32.6, the same as MTP alone |
| Qwen3.8 27B | Q4_K_M (15.3 GiB) | llama.cpp (CUDA, built 11 Nov 2026) | 25.0 | 1,118 @ 4K-token prompt est. TTFT @8k ≈ 7.3 s (calculated from measured prefill) | context: llama-bench, 4K, all on the RTX 5060 Ti📏 llama-bench pp4096 / tg128, 3 rounds under 1% apart (Somenzi, 11 Nov 2026); at 32K the model no longer fits: RTX + CPU 7.8, RTX + 780M layer split 11.8, tensor split 16.3 |
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-run series by Stefano Somenzi (10-12 Nov 2026). On the Radeon 780M alone (no eGPU, 16 Oct 2026, older Strata 0.1.44) the same Q2_0 generated 16.0 tok/s at 16K and prefilled 131 tok/s; with the RTX 5060 Ti it is about 6x faster on generation and 12x on prefill. PCIe Gen4 x4 is a quarter of a desktop slot, which slows long prompts on Strata because experts travel over the bus. Q2_0 is below Q4.
https://stefanosomenzi.com/llama-cpp-64k-context-16gb-qwen3-8-27b-speculative-decoding/
https://stefanosomenzi.com/llama-cpp-two-gpus-radeon-780m-rtx-5060-ti/
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Speculative-decoding rows on this page are not like-for-like with plain decode. See how much MTP and DFlash really speed things up (measured pairs).