Coding Horizon

Your 12GB GPU can run AI you might be overlooking

Every figure, price, date, version and benchmark this video puts on screen, chased to a primary source. Publisher pages are cited at the page the claim actually appears on, and the video shows that page rather than redrawing it wherever the claim is still theirs.

The machine

RTX 3060 12GB. 12GB GDDR6 on a 192 bit bus, about 360 GB/s of memory bandwidth. Desktop RTX 4070. Also a 12GB configuration, GDDR6X on a 192 bit bus, about 504 GB/s.

Equal capacity, different bandwidth: the 4070 moves roughly 1.4 times the bytes per second at the same 12GB. That is the whole of the script’s “equal capacity doesn’t mean equal speed” and it is why the figure on screen is bandwidth rather than capacity.

The arithmetic the script does

Nine billion weights at four bits is about 4.5 billion bytes. 9e9 weights x 4 bits = 36e9 bits = 4.5e9 bytes. The script states this as “a rough four bit calculation” and “before extra data and runtime overhead”, which is correct: a real GGUF carries embeddings, some tensors at higher precision, and the KV cache and context are on top of it again. The gap between that 4.5 and Ollama’s actual 6.6GB download is the honest illustration of it.

$1,000 shown, $200 refunded, $800 expected net. Arithmetic, and the fixture the whole video returns to.

40 tokens per second with 2,000 tokens of thinking is roughly 50 seconds. 2000/40 = 50. The script labels it “example numbers” in the same breath, and the shot carries that label.

Qwen 3.5 9B

Alibaba’s model, 9B parameters, text and image in, 256K context as Ollama lists it.

claim value where
Ollama download size 6.6GB for qwen3.5:9b Ollama library page
LiveCodeBench v6 65.6 Qwen’s own model card table
MMLU Pro 82.5 same table
GPT OSS 20B, MMLU Pro 74.8 same table, comparison column
GPT OSS 20B, LiveCodeBench v6 74.6 same table, and it LEADS qwen there

That last row is the script’s “In that same table, GPT OSS leads it on livecodebench”, and it is the reason the beat draws both rows of the same table rather than only the one qwen quotes.

One contradiction found and resolved. A secondary write up (XDA) reports LiveCodeBench v6 of 82.7 for this model. Qwen’s own card says 65.6. The card is the primary source and 65.6 is what goes on screen; the 82.7 figure is not used anywhere.

Gemma 4 12B

Google’s model. Encoder free, so text, image AND audio go into one decoder.

claim value where
LiveCodeBench v6 72.0 Google’s Gemma 4 model card
Audio input on the 12B yes, up to 30s same card
Four bit GGUF size 7.22GB, gemma-4-12B-it-Q4_0.gguf ggml-org repo
Separate vision file mmproj, 0.16 to 0.18GB same repo
Visual token budget “A higher token budget preserves more visual detail at the cost of additional compute” and higher budgets are recommended “for tasks like OCR, document parsing, or reading small text” same card

A note on which repo. ggml-org/gemma-4-12B-GGUF (the base) carries only BF16 and Q8_0. The four bit file is in ggml-org/gemma-4-12B-it-GGUF, the instruction tuned repo, which is the one anybody running a coding assistant would pull. The script’s “about 7.2GB” is 7.22GB exactly.

Ministral 3 14B Reasoning

Mistral’s model, in the 2512 release.

claim value where
Bartowski four bit file 8.24GB, ...Q4_K_M.gguf bartowski’s GGUF repo
Separate vision component mmproj, 0.88 to 0.89GB same repo
LiveCodeBench 64.6 Ministral 3 paper, Table 5
Qwen 3 14B comparison 59.3 same table, same row
Protocol pass@16 throughout, except LiveCodeBench at pass@5 same paper

Two things the shot has to be careful about. Table 5’s 14B column is headed Qwen 3 14B, which is the Qwen 3 generation and not Qwen 3.5, and the script says exactly that (“That’s the older qwen 3, not qwen 3.5”). The paper’s 8B and 4B columns are headed Qwen3-VL; only the 14B pair is the one this video draws. And the score belongs to the REASONING model: Mistral ships a separate instruct variant, so the benchmark is not a score for every file carrying the ministral name, which is the point the beat makes.

Ternary Bonsai 2 27B

PrismML’s compression of Qwen3.8 27B. Released 17 September 2026, Apache 2.0.

claim value where
Release date 17 September 2026 PrismML launch post
Retention 98.2% of FP16 aggregate, 84.78 against 86.32 across 14 thinking mode benchmarks model card and launch post
Measured against Qwen3.8 27B full precision both
Packings 5.95GB (TQ1_0, 1.75 bpw) and 7.21GB (PQ2_0, 2.13 bpw) model card
Separate vision file mmproj, 0.63GB Q8_0, optional model card
Peak throughput up to 143 tok/s on an RTX 5090 launch post
Runtime the PrismML-Eng llama.cpp fork; “stock llama.cpp will not run these files” model card

The compression comparison, which is the chapter’s real point

The model card’s own table, all four variants of the same parent:

Variant bpw Footprint Thinking avg vs FP16 LiveCodeBench
Qwen3.8-27B FP16 16.0 54 GB 86.32 100% 90.05
Qwen3.8-27B UD-Q4_K_XL 5.2 17.6 GB 85.18 98.7% 87.96
Qwen3.8-27B IQ2_XXS 2.8 9.4 GB 72.59 84.1% 56.40
Ternary Bonsai 2 27B 1.72 5.9 GB 84.78 98.2% 90.07

That is the script’s “bonsai scores about 90 on livecodebench, while a conventional, very low bit version of the same parent scores about 56”, exactly, and it is two compressions of ONE model behaving completely differently at a similar bit width.

The version caveat is real and the script is right to make it. This table’s row is labelled only LiveCodeBench, with no version. Google’s 72 and Qwen’s 65.6 are both explicitly v6. So these numbers cannot be dropped into the same chart, and the video does not draw them on one axis.

Ollama’s context default

Ollama picks a default from available VRAM: 4096 tokens below 24GiB, 32768 from 24 to 48GiB, 262144 at 48GiB and above. Any 12GB card is in the first tier.

A documentation inconsistency worth knowing. Ollama’s FAQ says 4096, the Modelfile reference says num_ctx defaults to 2048, and the context length page describes the VRAM tiers. The context length page is the current and most specific one, and it is what the video cites and shows.

Derived on screen, and where from

Not chased to a primary source