Builds / NVIDIA T1000 4GB
NVIDIA T1000 4 GB + i7-12700K: Qwen3 30B-A3B IQ3_XXS at 18.7 tok/s with experts in RAM, 27.1 with a 0.6B draft model
Last updated:
Source: github.com ↗More source: github.com ↗📏 How measured: reported by the original source; RigBase re-checked the numbers at that source
Legacy models only. Every model measured on this rig now has a newer successor: . The hardware numbers are still a useful reference.
- GPU(s)
- NVIDIA T1000 4GB
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 4
- CPU
- Intel Core i7-12700K (8P+4E)
Amazon · eBay (used) · Newegg - Motherboard
- not stated
Amazon · eBay (used) · Newegg - RAM
- 32 GB
Amazon · eBay (used) · Newegg - PSU
- not stated
- Other
- Windows, llama-server b12121+; -ngl 99 -ncmoe 41 (experts of 41 of 48 layers in system RAM), --load-mode none, -fa on, q8_0 KV, 8K context, P-core affinity (-Cr 0-15); 3,762 of 4,096 MiB VRAM used
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Qwen3 30B-A3B legacy | Unsloth UD-IQ3_XXS (12.0 GB) | llama.cpp llama-server b12121+ (CUDA) | 18.68 | not stated | context: 8K ctx, all 8 experts per token📏 Static expert offload (-ncmoe 41), full 8-expert routing, no draft: 18.68 tok/s (baseline for the draft row) |
| Qwen3 30B-A3B legacy | Unsloth UD-IQ3_XXS (12.0 GB) | llama.cpp llama-server b12121+ (CUDA) | 27.11 | not stated | ⚡ spec-decode (draft model (Qwen3 0.6B UD-Q4_K_XL))context: 8K ctx, all 8 experts per token📏 Same setup plus --spec-type draft-simple --spec-draft-n-max 3 --spec-draft-p-min 0.80: 27.11 tok/s, 94.8% acceptance (91/96 tokens) |
Owner's 6-stage study posted as a llama.cpp discussion (10 Nov 2026). The headline 39.18 tok/s in the post uses only 4 of 8 experts per token (expert_used_count=4), which changes the model's output quality, so it is not listed as a result here. A dynamic GPU expert cache (--moe-cache-mib 768) was much slower (8.57) than static pinning on this 4 GB card. IQ3_XXS is below Q4.
Source: https://github.com/ggml-org/llama.cpp/discussions/32007
https://github.com/AjayakumarV/Qwen3Studio-LocalAI
https://github.com/AjayakumarV/Qwen3Studio-LocalAI
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)
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).