Coding Horizon

Your $1,000 local AI budget can buy the wrong machine

Sources for every figure this video puts on screen. Prices are US dollars and exclude sales tax and shipping. A complete machine budget assumes an existing display, keyboard and mouse. Asking prices, launch prices and published file sizes are different kinds of evidence, and availability is never inferred from an indexed price.

Checked 16 and 17 September 2026.

Graphics cards

On screen	Figure	Source
$780 card	Newegg lists MSI Shadow RTX 5060 Ti 16GB at $779.99 and Ventus 16GB at $799.99	Newegg RTX 5060 Ti 16GB listings
$429 launch price	NVIDIA announced the RTX 5060 Ti 16GB at $429 at launch	NVIDIA Blackwell GeForce RTX pricing release
16 GB on the card	NVIDIA specifies 16GB on the RTX 5060 Ti 16GB	NVIDIA RTX 5060 Ti specifications
RTX 3090, 24 GB, about $1,500	NVIDIA specifies 24GB. eBay shows used asking prices of $1,499 and $1,500	NVIDIA RTX 3090 specifications, eBay RTX 3090 listings
RTX 3060, 12 GB	NVIDIA specifies both 12GB and 8GB variants of the RTX 3060	NVIDIA RTX 3060 family specifications
Intel Arc B580, 12 GB, about $300	Intel specifies 12GB; Micro Center lists a Sparkle B580 at $299.99	Intel Arc B series launch, Micro Center listing
Radeon RX 9060 XT 16 GB, about $500	Newegg lists a Sapphire Pulse at $499.99 and an ASRock Challenger at $519.99	Newegg RX 9060 XT 16GB listings

Complete machines

On screen	Figure	Source
Legion T5, $799, sold	OfferUp listing: i7-11700F, RTX 3060 12GB, 16GB DDR4, $799, marked SOLD	Sold Lenovo Legion T5 listing
$700 to $800 hunting range	Two further sold RTX 3060 towers at $799 and $699, both with 32GB and a 1TB SSD	Spokane listing, Glendale listing
M6 mac mini, $899, 16 GB, September 22	Apple announced the M6 mini on 25 August: $899 starting price, 16GB standard memory, preorder with deliveries from 22 September	Apple Mac mini announcement
Up to 4.8 times faster	Apple reports up to 4.8x faster LLM prompt processing versus M4 in its own test. That is reading the request, not finishing a coding task	Apple Mac mini announcement
Used M4 mini with 24 GB	Apple documents 16GB base with 24GB and 32GB configurations	Apple M4 mini specifications
GMKtec K8 Plus, $400 barebone	Manufacturer page: $399.99 selected barebone, no RAM, storage or OS. Ryzen 7 8845HS, Radeon 780M, two DDR5 SO-DIMM slots	GMKtec K8 Plus
EVO X1 Pro, $1,600	Official US catalog: $1,599.99 starting price, Ryzen AI 9 HX 470	GMKtec EVO X1 Pro
EVO X2, above $2,000	Official US catalog: $2,199.99 starting price, Ryzen AI Max+ 395	GMKtec EVO catalog

The $1,000 figure, the $900 reserve for a K8 Plus build and the $900 ceiling on a used M4 mini are the video’s own editorial limits, not prices quoted by anyone.

Models and file sizes

Sizes are the decimal GB the quantization publisher displays. They are weights on disk and exclude the KV cache, runtime buffers and any separate vision projector. Hardware capacity is often quoted in binary units, so 16 GiB is about 17.18 decimal GB and a 16.5 GB file does not by itself prove the weights exceed a 16 GB card. The constraint the video states is the remaining headroom, not the arithmetic.

On screen	Figure	Source
Qwen 3.5 9B, nine billion	Official Qwen model, 9B dense	Qwen3.5 9B model card
17.9 GB and 5.68 GB	Unsloth publishes BF16 at 17.9 GB and Q4_K_M at 5.68 GB	Qwen3.5 9B GGUF listing
Ministral 3, 8B and 14B, 8.24 GB	Official Mistral compact instruct family; Unsloth 14B Q4_K_M is 8.24 GB	Ministral 3 14B, GGUF listing
DeepSeek R1 distilled Qwen 14B	Official DeepSeek distillation built on Qwen, and not the DeepSeek service	DeepSeek R1 Distill Qwen 14B
Qwen 3.8 27B, released august, 16.5 GB	August 2026 release, dense 27B; Unsloth UD-Q4_K_M is 16.5 GB	Qwen3.8 27B model card, GGUF listing
Devstral Small 2, 24 billion, 14.3 GB	Mistral software engineering model; Unsloth Q4_K_M is 14.3 GB	Devstral Small 2, GGUF listing
Qwen3 Coder 30B, 18.6 GB	Coding and tool use model; Unsloth Q4_K_M is 18.6 GB	Qwen3 Coder 30B GGUF listing
Qwen3 Coder Next, 80 billion total, 3 billion active	Official card: 80B total, 3B active per token	Qwen3 Coder Next
DeepSeek V4 Flash, 284 billion	Official card: 284B total	DeepSeek V4 Flash
Kimi K3, 2.8 trillion	Official card: 2.8T total	Kimi K3
61.7% and 68.0%	Qwen reports 61.7 on SWE bench Pro for Qwen3.8 27B; Mistral reports 68.0% on SWE bench Verified for Devstral Small 2. Different benchmarks and setups, so the two do not rank against each other	Qwen and Mistral published evaluations on the model cards above

Memory, bandwidth and runtime support

• A model’s weights are only part of what has to be resident. Working buffers and the conversation’s KV cache are additional, and the cache grows with the amount of code held in context. llama.cpp documents the quantization formats, the backends and hybrid CPU and GPU operation this depends on.
• On a discrete card, VRAM is dedicated to that card. On Apple silicon the CPU and GPU share one unified pool with the operating system and running applications, so installed memory is not memory available to a model. Apple M4 mini specifications
• Shared DDR capacity and dedicated graphics bandwidth solve different constraints. Capacity lets a larger model load; bandwidth is what each round trip pays.
• Ollama’s matrix lists the RTX 3060, 3090 and 5060 Ti. Its AMD ROCm list on linux includes the RX 9060 XT and Ryzen AI HX 470; the Radeon 780M is absent from the standard listed devices, and the windows list differs. Absence from the matrix does not prove an override or a different runner cannot work. Ollama GPU support matrix
• Intel documents a llama.cpp installation route for B series cards. Intel llama.cpp guide
• MLX provides quantized inference on Apple silicon alongside Metal. MLX LM

Screenshots used on screen

One Hugging Face file listing appears as a captured page rather than a redrawn figure, so the size on screen is the publisher’s own:

• unsloth/Qwen3.8-27B-GGUF, files and versions

Not established here

• No tokens per second, latency or throughput result is claimed for any machine. The fit judgements are estimates from published model sizes, memory capacity and documented runtime support.
• The timing comparison late in the video is a method to run, not a measurement that was taken. The two bars illustrate how total time to a passing test can invert a per reply speed advantage; they are not results.
• Retailer and marketplace listings are prices displayed on the dates checked, not guaranteed stock and not the minimum obtainable price. Sold listings are asking prices, not confirmed transaction prices.
• The memory splits drawn in the capacity beats are representative allocations for a desktop with an editor and tests open, not measurements of a particular machine.
• Ollama listing a device means the runner supports it, not that every model runs well on it.