Coding Horizon

A $300 PC can run 27B AI, but there’s a huge catch

Every figure, price, capacity and benchmark the picture puts on screen, chased to a primary source. Checked 23 September 2026. US listings, prices before sales tax unless stated.

Three claims could not be chased to a primary source and are listed under Not sourced at the foot. Nothing in that section reaches the screen as a figure.


The budget arithmetic

The video works to an example eight percent sales tax. Every figure it states is the arithmetic on the listed price, and all of it holds:

Stated Arithmetic Result
a $300 ceiling becomes about $277 before tax 300 / 1.08 $277.78
$276.35 comes to $298.46 276.35 x 1.08 $298.46
the $330 machine becomes roughly $356 329.99 x 1.08 $356.39
$329.99 is $30 over budget 329.99 - 300 $29.99
$275 before tax, as a ceiling for a strict $300 purchase 275 x 1.08 $297.00
32 GB is four times the 8 GB mini PC 32 / 8 4
16 GB of RAM plus 6 GB of VRAM would be 22 GB 16 + 6 22

The eight percent rate is the video’s own worked example, not a rate that applies anywhere in particular. Actual US sales tax is set by state and locality.

The machines

GMKtec G3S, Intel Alder Lake N95, $249.99. The manufacturer’s own US store lists the G3S at $249.99. The N95 is a four core Alder Lake N part with integrated Intel UHD graphics and no dedicated video memory. https://www.gmktec.com/collections/mini-pc

GMKtec G10, AMD Ryzen 5 3500U, advertised from $199.99. The manufacturer’s product page lists the G10 from $199.99 and offers three configurations: Barebone (No DDR & SSD & OS), 16GB + 512GB SSD and 16GB + 1TB SSD. The page states outright that “The Barebone model comes without DDR memory, SSD, and OS. You need to install your own compatible DDR memory and SSD before powering on the device.” The $199.99 headline is therefore the barebones price, and the complete computer is a different SKU. https://www.gmktec.com/products/gmktec-g10-amd-ryzen-5-3500u-mini-pc

The complete 16 GB / 512 GB configuration price is served by the store’s variant selector rather than as page text, so the exact $329.99 could not be read off a static capture. See Not sourced.

Dell OptiPlex 7060 small form factor: the upgrade limit is real. Dell ships the SFF chassis with a 200W power supply and a single low profile PCIe x16 slot. A full height, dual slot GTX 1660 does not fit that slot and is not inside that power budget, so it is not a drop in upgrade. https://www.dell.com/support/manuals/en-us/optiplex-7060-sff/opti_7060_sff_setup_specs_manual/power-supply https://www.hardware-corner.net/desktop-models/Dell-OptiPlex-7060-SFF/

The i5 8500 is a six core, six thread Coffee Lake desktop part, which is the “six core desktop CPU” the video compares against the N95’s four cores.

The graphics cards

GTX 1660 and 1660 Super carry 6 GB of GDDR6. The GTX 1650 Super carries 4 GB. The video’s warning that similar names mean meaningfully different room for a model is the 50% capacity difference between those two parts. https://www.techpowerup.com/gpu-specs/geforce-gtx-1660-super.c3458 https://www.techpowerup.com/gpu-specs/geforce-gtx-1650-super.c3463

The 16 series has no tensor cores. The GTX 16 family is built on Turing TU116 and TU117, which are the Turing dies with the tensor and RT cores removed. They keep full CUDA compute capability 7.5, so CUDA inference software still runs on them; what they do not have is the tensor hardware the RTX cards use.

The models

Qwen 3.5 download sizes, from Ollama’s own library page. These are the figures the video quotes and they are exact:

Tag Download
qwen3.5:2b 2.7 GB
qwen3.5:4b 3.4 GB
qwen3.5:9b 6.6 GB
qwen3.5:latest 6.6 GB, and it is the 9b

https://ollama.com/library/qwen3.5

The last row is the video’s point about default tags: ollama run qwen3.5 with no tag pulls the 9B, not the small model.

Qwen 3.5 LiveCodeBench v6: 4B scores 55.8, 9B scores 65.6. Both figures are from Qwen’s own published model card, in the row labelled LiveCodeBench v6. The separation is 9.8 points, which is the “roughly ten points” the video states. https://huggingface.co/Qwen/Qwen3.5-9B

Gemma 4 memory requirements, from Google’s own table. The page is headed Gemma 4 Inference Memory Requirements and carries its own caveat that “These numbers may change based on your specific inference tool and environment.”

Model BF16 (16-bit) SFP8 (8-bit) Q4_0 (4-bit)
Gemma 4 E2B 11.4 GB 5.7 GB 2.9 GB
Gemma 4 E4B 17.9 GB 8.9 GB 4.5 GB

https://ai.google.dev/gemma/docs/core

The default Ollama gemma4 packages are much larger than the official four bit files. gemma4:e2b is 7.2 GB and gemma4:e4b is 9.6 GB, against the 2.9 GB and 4.5 GB above. That gap is the video’s point about matching the model, the quantization and the file format together. https://ollama.com/library/gemma4

Ternary Bonsai 2 27B, from PrismML. Every file size the video states is exact, and comes from the repository’s own file listing:

File Size
Language model, PTQ1_0 5.95 GB
Language model, PQ2_0 7.21 GB
Image projector, mmproj Q8_0 0.63 GB
F16 reference 53.8 GB

It is built on Qwen3.8 27B, is released under Apache 2.0, and requires PrismML’s own llama.cpp fork: the repository states that stock llama.cpp “rejects PQ2_0 and PTQ1_0 as unknown types”. CUDA, Metal and CPU backends are supported, and a separate MLX build exists for Apple Silicon. https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf https://prismml.com/news/bonsai-2-27b

The video’s caution that a 5.95 GB file is not a promise it runs inside 6 GB of VRAM is the ordinary difference between weights on disk and weights plus KV cache, activation buffers and runtime overhead in memory. PrismML’s own guidance is a 16 GB laptop or a single 24 GB GPU.

The marketplaces

Jawa: a GTX 1660 system with 16 GB and a 512 GB SSD at $269.99, sold. The listing is “GTX 1660 | Intel i5 | 16GB DDR4 RAM | 512GB SSD | DVD Slot | Gaming PC” from the seller Attack Computers, priced at $269.99 down from $299.99, published 23 November 2025 and now sold out. That is exactly the shape of the video’s claim: evidence of a past price, not a checkout button. https://www.jawa.gg/product/90740/black-friday-gtx-1660-or-intel-i5-or-16gb-ddr4-ram-or-512gb-ssd-or-dvd-slot-or-gaming-pc

Swappa: $299 M1 Mac mini, 8 GB, 256 GB. A Mac mini 2020 with Apple M1, 8 GB and 256 GB at $299 in mint condition is listed on Swappa. Live Swappa listings at this configuration span roughly $265 to $478 depending on condition. https://swappa.com/listing/view/LYHH77291

The M1 Mac mini shares one pool of memory between CPU and GPU, and local runtimes reach its GPU through Apple’s Metal API. The operating system draws on that same pool, which is the video’s point that an 8 GB machine does not hand a model eight empty gigabytes. The memory is soldered, so 8 GB is a permanent ceiling on that unit.

Activation Lock and remote management are real blockers on a used Mac. A Mac still signed in to a previous owner’s Apple Account, or still enrolled in an organisation’s mobile device management, will ask for those credentials during setup and cannot be cleared by the buyer.

Not sourced

Three figures the narration states could not be chased to a primary source. None of them is rendered as a figure on screen.

All three are marketplace prices, and all three carry the video’s own caveat with them: a sold listing is evidence of a past price rather than inventory anyone can order today.