Fine Tune 8B AI On A $480 RTX 3060 (One Catch)
Sources for every figure on screen, checked on 2026-10-10.
The Hugging Face demo
- Qwen3-0.6B on open-r1/codeforces-cots, SFT. Hugging Face’s post shows the agent proposing
“fine-tune Qwen/Qwen3-0.6B on open-r1/codeforces-cots using SFT”.
- t4-small, ~$0.75/hour, ~20 minutes, ~$0.30. The configuration block in the same post: “Hardware:
t4-small (~$0.75/hour)”, “Estimated time: ~20 minutes”, “Estimated cost: ~$0.30”. All three are marked
as estimates. The post gives no row count for the dataset and no production run’s GPU hours.
- A 100 example test run. The post suggests “Do a quick test run on 100 examples” before a full job.
- Jobs need a paid plan. The post: “Jobs require a paid plan.”
- Budget ranges by model size. The post’s hardware guide: under 1B “expect $1-2 for a full run”;
1-3B “costs $5-15”; 3-7B “Budget $15-40 for production”.
https://huggingface.co/blog/hf-skills-training
Hugging Face Jobs prices (per hour, billed by the minute)
- Nvidia A100 large (80 GB): $2.50. Nvidia H200 (141 GB): $5.00. Nvidia RTX PRO 6000 (96 GB): $2.75.
- The same table lists T4 small at $0.40 an hour today; the video quotes the tutorial’s ~$0.75.
https://huggingface.co/docs/hub/main/en/jobs-pricing
Minimum VRAM for fine tuning (Unsloth)
| model |
QLoRA (4-bit) |
LoRA (16-bit) |
| 3B |
3.5 GB |
8 GB |
| 7B |
5 GB |
19 GB |
| 8B |
6 GB |
22 GB |
| 9B |
6.5 GB |
24 GB |
| 11B |
7.5 GB |
29 GB |
| 14B |
8.5 GB |
33 GB |
| 32B |
26 GB |
76 GB |
Unsloth calls these the “absolute minimum”, notes more VRAM is sometimes needed depending on the
model, and recommends a batch size of 1, 2 or 3 when memory runs out.
https://unsloth.ai/docs/get-started/fine-tuning-for-beginners/unsloth-requirements
Unsloth’s Intel GPU guide fine tunes Qwen3-32B in 4-bit (unsloth/Qwen3-32B-bnb-4bit) on Intel GPUs
(Data Center GPU Max, Arc series, Core Ultra AI PCs).
https://unsloth.ai/docs/get-started/install/intel
Hardware prices, checked 2026-10-10
- RTX 3060 12GB, $479.99. MSI RTX 3060 Ventus 2X 12G OC, sold by Newegg. The cheapest Newegg-sold
RTX 3060 12GB that day was a Gigabyte Windforce OC at $459.99.
https://www.newegg.com/p/pl?d=rtx+3060+12gb&N=8000%204131&Order=1
- Used RTX 3090 24GB, about $1,600. RigPrice median asking price: $1,600 on its 2026-10-07 snapshot,
$1,619 on 2026-10-10 (77 active used listings). https://rigprice.com/gpu/rtx-3090/
- Intel Arc Pro B60 24GB, $649.99. ASRock B60 CT 24G, sold and shipped by Newegg.
https://www.newegg.com/p/N82E16814930148
- Framework Desktop, Ryzen AI Max+ 395, 128GB: $3,449 (DIY edition, before storage and OS; out of
stock on 2026-10-10). https://frame.work/products/desktop-diy-amd-aimax300/body/new
- Mac Studio. M5 Max from $2,499, configurable to 128GB; M5 Ultra from $5,499, up to 512GB (the 512GB
configuration arrives late October). Apple names MLX for fine tuning on a Mac.
https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/
- AMD Ryzen AI Max+ 395 specifications:
https://www.amd.com/en/products/processors/desktops/ryzen/ryzen-ai-halo/ryzen-ai-max-plus-395.html
Fine tuning on Strix Halo and Mac
- kyuz0’s Strix Halo fine tuning toolbox (Fedora toolbox, ROCm) has notebooks for full fine tuning,
LoRA, 8-bit LoRA and QLoRA on Gemma 3 and Qwen 3, with memory use and wall-clock time per run. It
reports no tokens per second, so its times cannot be set against a rented GPU’s.
https://github.com/kyuz0/amd-strix-halo-llm-finetuning
- MLX LM supports LoRA and QLoRA fine tuning on Apple silicon. https://github.com/ml-explore/mlx-lm
- Apple, WWDC 2026 session 233, MLX training and fine tuning across devices.
https://developer.apple.com/videos/play/wwdc2026/233/
| figure |
working |
| about 38 cents |
half an hour at $0.75 = 0.375 |
| $1.25 |
half an hour at $2.50 |
| 32 to 96 jobs |
480 / 15 = 32; 480 / 5 = 96 |
| about 16 runs |
650 / 40 = 16.25 |
| 40 runs |
1,600 / 40 |
| about 43 and 107 runs |
650 / 15 = 43.3; 1,600 / 15 = 106.7 |
| $10 to $11 |
4 hours at $2.50 and at $2.75 |
| $120 a year |
12 × $10 |
| over 5 years |
650 / 120 = 5.4 |
| over 13 years |
1,600 / 120 = 13.3 |
| 260 hours |
650 / 2.50 |
| 640 hours |
1,600 / 2.50 |
| 2 GB short |
26 − 24 |