Every figure the video shows or says, with where it came from. Prices were checked on 14 October 2026 (the script’s October 10 figures were re-checked that day; the ones that moved were corrected) and move often; check the live listing before buying.
Source: Hugging Face, prism-ml/Ternary-Bonsai-2-27B-gguf, read 14 October 2026.
| Figure | Source text |
|---|---|
| PTQ1_0 file 5.95 GB; PQ2_0 file 7.21 GB | “PTQ1_0 packs trits densely (1.75 bits/weight, 5.95 GB), PQ2_0 stores each trit in a 2-bit slot (2.13 bits/weight, 7.21 GB)” |
| Vision projector 0.63 GB, optional | “Vision tower mmproj (Q8_0) 0.63 GB optional — multimodal input only” |
| 262,144-token context ceiling | “262K-token context”; “-c sets the context, up to 262144” |
| Stock llama.cpp cannot run the files | “Stock llama.cpp will not run these files. It rejects PQ2_0 and PTQ1_0 as unknown types, and it loads Q2_0 without any warning and produces garbage” |
| M5 Pro laptop, 28.1 tok/s, PQ2_0 | Cross-Platform Throughput table: “Laptop (Apple M5 Pro, Metal) 28.1” (tg128, batch 1, depth 0, no vision tower) |
| RTX 5090, 129.9 tok/s, PQ2_0 | Same table: “RTX 5090 (32 GB) 129.9” (“measured with llama-bench on these GGUF packs”) |
| 14-benchmark average: Bonsai 2 84.78, FP16 Qwen3.8-27B 86.32 (98.2%) | Benchmarks table: “Qwen3.8-27B FP16 … 86.32 100%”, “Bonsai 2 27B … 5.95 GB 84.78 98.2%” |
| 2-bit Qwen (IQ2_XXS): 7.27 GB, 72.59 | “Qwen3.8-27B IQ2_XXS (“2-bit”) 2.16 7.27 GB 72.59 84.1%” |
| LiveCodeBench: 90.07 Bonsai 2, 90.05 FP16 | Full per-benchmark results: “LiveCodeBench 90.05 87.96 56.40 90.07” |
| OCRBench v2: 56.88 Bonsai 2, 65.45 four-bit Qwen (UD-Q4_K_XL) | “OCR Bench v2 60.99 65.45 61.70 56.88” |
No RTX 5060, RTX 3060, RTX 5060 Ti, RTX 3090 or Mac mini row exists on the model card’s throughput table.
Source: PrismML-Eng/Bonsai-demo README.
| Claim | Source text |
|---|---|
| The demo downloads PQ2_0 by default | “A second packing, PQ2_0, trades 1.3 GB for faster prompt processing and is what this demo downloads by default” |
| Needs the PrismML llama.cpp fork | “Bonsai 2 needs this demo’s llama.cpp binaries, from the PrismML fork; stock llama.cpp cannot run these files” |
| Vision projector can stay in system RAM | “BONSAI_MMPROJ_CPU=1 keeps the 27B’s vision projector in system RAM instead of VRAM … The cost is a slower image prompt” |
| Apple Silicon: MLX build | Model card: “MLX companion: also available as Ternary-Bonsai-2-27B-mlx-2bit for native Apple Silicon inference” |
| Figure | What it is | Source |
|---|---|---|
| $460 | Cheapest new RTX 5060 8GB (Gigabyte Windforce Max OC V2, $459.99) | Newegg search, lowest price first |
| $350 | RTX 3060 12GB, cheapest screened used pick | RigPrice RTX 3060 12GB (“Cheapest screened pick $350 used”) |
| about $370 | RTX 3060 12GB used going rate ($375 on the day) | RigPrice |
| $460 | RTX 3060 12GB new (Gigabyte Windforce OC rev 2.0, “Buy New $459.99”) | Newegg |
| $800 | MSI RTX 5060 Ti 16G Gaming Trio OC, sold by Newegg ($799.99) | Newegg |
| $1,590 | RTX 3090 24GB, cheapest screened pick, refurbished | RigPrice RTX 3090 |
| $1,600 | RTX 3090 used-condition listing (EVGA FTW3 Ultra, $1,599) | RigPrice |
| about $1,649 | RTX 3090 used going rate | RigPrice |
| $6,886 | Gigabyte RTX 5090 Windforce OC 32G, sold by a marketplace seller (Slava computers), condition not stated ($6,885.90) | Newegg |
| $899 | Mac mini M6, 16GB, 256GB | Apple |
| $1,699 | Mac mini M5 Pro, 24GB, 512GB | Apple |
| $2,699 | The same Mac mini with the 64GB option (+$1,000) | Apple configurator, our arithmetic: $1,699 + $1,000 |