Coding Horizon

Mac mini vs RTX 5090: don’t buy the wrong AI machine

Sources for every figure, price, date and benchmark the picture puts on screen.

The script was written and recorded before the shots existed, so a claim the narration makes cannot be corrected by changing a shot. Anything below that could not be chased to a primary source is not rendered on screen at all, and is listed under “Not put on screen” at the foot of this file.

Every figure the picture renders, through all seventy one beats, is below. Beat numbers in this file are this cut’s, and the SHOT: lines in BEATS.md are the index back into it.

Mac mini, M5 Pro and M4 Pro

Fact Value Source
M5 Pro Mac mini starting price $1,699 Apple Newsroom, 25 August 2026
M5 Pro standard unified memory 24GB Apple, Mac mini technical specifications
M5 Pro maximum unified memory 64GB Apple, Mac mini technical specifications
M5 Pro memory bandwidth 307GB/s Apple, Mac mini technical specifications
M4 Pro memory bandwidth 273GB/s Apple, Mac mini (2024) technical specifications
M4 Pro maximum unified memory 64GB Apple, Mac mini (2024) technical specifications
M6 Mac mini starting price $899 Apple Newsroom, 25 August 2026
M6 maximum unified memory 32GB Apple, Mac mini technical specifications
Mac mini maximum continuous power 155W Apple, Mac mini technical specifications

Memory is part of the system on a chip package and is configured at order. Apple lists no user accessible memory slot for any Mac mini configuration, which is what beats 010 and 055 draw as a cutaway rather than assert as a sentence.

The M4 Pro is the load bearing one in chapter 6 (beats 053 and 054): it holds the same 64GB as the M5 Pro and moves it at 273GB/s against 307, which is the 11 per cent the narration calls “a little less speed”. Both figures are Apple’s own.

GeForce RTX 5090

Fact Value Source
Launch price $1,999 MSRP, January 2025, card only NVIDIA, GeForce RTX 5090
Memory 32GB GDDR7 on a 512 bit bus NVIDIA, GeForce RTX 5090
Memory bandwidth 1,792GB/s NVIDIA, GeForce RTX 5090
Total graphics power 575W NVIDIA, GeForce RTX 5090
Recommended system power supply 1000W NVIDIA, GeForce RTX 5090

575W is the board’s total graphics power, which is a peak rating rather than the draw of any particular workload. Nothing in the sources states a per query figure, so beat 061 draws the rating as a ceiling with a varying trace under it and prints no inference watts.

The 1,792 against 307 comparison in beat 045 needs no separate sourcing: both numbers are in the two tables above, from NVIDIA’s and Apple’s own specifications. The multiple between them, 5.8, is computed from the pair in the shot rather than typed, and beat 046 strikes it out as a speed claim while leaving the bandwidth figure standing.

Qwen3.8-27B

Fact Value Source
Parameters 27 billion Qwen/Qwen3.8-27B model card
SWE bench Pro 61.7 Qwen/Qwen3.8-27B model card
SWE bench Pro, Qwen3.6-27B, same card 53.5 Qwen/Qwen3.8-27B model card
Vision native vision language model, understands images and video Qwen/Qwen3.8-27B model card
Context length 262,144 native, extensible to 1,000,000 Qwen/Qwen3.8-27B model card

Both figures are the developer’s own published evaluations, run with an agent harness and a large context window. They describe that setup, not a four bit download on a desk, which is what beat 008 labels in amber.

The llama.cpp team’s GGUF conversion

File Size Source
Qwen3.8-27B-Q4_K_M.gguf 19GB ggml-org/Qwen3.8-27B-GGUF
Qwen3.8-27B-Q8_0.gguf 28.6GB ggml-org/Qwen3.8-27B-GGUF
Qwen3.8-27B-BF16.gguf 53.8GB ggml-org/Qwen3.8-27B-GGUF

ggml-org is the organisation the llama.cpp maintainers publish under. The 19GB figure is the weights file only. Conversation history, image tokens and intermediate state are additional and runner dependent, which is why beat 013 draws the parts and never sums them.

Mistral Devstral Small 2, and Devstral 2

Fact Value Source
Devstral Small 2 parameters 24,011,361,840, the 24 billion the narration says mistralai/Devstral-Small-2-24B-Instruct-2512
Devstral Small 2, SWE bench Verified 68.0% the same model card’s benchmark table
Devstral Small 2 context length 256k the same model card
Devstral 2 parameters 123B, dense Mistral AI, Introducing Devstral 2 and Mistral Vibe CLI
Devstral 2, SWE bench Verified 72.2% Mistral AI, Introducing Devstral 2 and Mistral Vibe CLI
Devstral 2, SWE bench Multilingual 61.3% the same announcement
Devstral 2 context length, licence 256K, modified MIT mistralai/Devstral-2-123B-Instruct-2512

Both Devstral figures are on SWE bench Verified, so beat 017 may put them on one axis. That is the opposite of beat 008’s situation and it is why the two beats are drawn differently: 68.0 against 72.2 is one ruler, and 61.7 on SWE bench Pro is another.

Five times the parameters for 4.2 points is arithmetic on the pair (123 / 24 = 5.1; 72.2 minus 68.0 = 4.2), computed in beat 018 from the two sourced figures rather than typed.

Google Gemma 4 12B

Fact Value Source
Parameters 11,959,730,224 google/gemma-4-12B-it
Input modalities text, image and audio the same model card
Context length 256K the same model card

Audio is native on the 12B rather than an add on, which is what makes the narration’s spoken bug report in beats 036 and 037 a real example rather than a hypothetical one, and what beat 038 is contrasting when a larger model drops the feature.

z.AI GLM 4.7 Flash

Fact Value Source
Total parameters 30B, a 30B A3B mixture of experts zai-org/GLM-4.7-Flash
Activated per token 3B the same model card
SWE bench Verified 59.2 the model card’s own comparison table
Qwen3-30B-A3B-Thinking-2507, same table 22.0 the model card’s own comparison table
Routing 64 routed experts plus 1 shared the same repository’s config.json
Licence MIT the same model card

59.2 against 22.0 is z.AI’s own published comparison of its model against somebody else’s, which the narration says out loud (“in z.AI’s evaluation”) and beat 022 labels on screen. Both models are about thirty billion parameters, which is the beat’s whole point, so the two rows carry the same size qualifier.

The 3 billion active against 30 billion held is what beats 021 and 023 draw as a lit fraction of a grid: cheaper to run, and no cheaper to hold.

Qwen3 Coder Next

Fact Value Source
Total parameters 79,674,391,296, stated as 80B on the card Qwen/Qwen3-Coder-Next
Activated per token 3B the same model card
SWE bench Verified 70.6 the same model card’s evaluation table
SWE bench Pro 44.3 the same model card’s evaluation table
Terminalbench 2 36.2 the same model card’s evaluation table
Four bit MLX conversion 44.9GB across nine safetensors shards mlx-community/Qwen3-Coder-Next-4bit
Context length 256k the same model card

70.6 and Devstral Small 2’s 68.0 are both on SWE bench Verified, so the gap is real and comparable as a number. What is not comparable is the harness each vendor ran it under, and that is the distinction beat 031 makes: the bars are drawn on one axis, and the qualifier under each names whose evaluation it is.

44.9GB is the weights only, which is what beat 028 puts inside the 64GB rail and beat 029 runs off the end of the 32GB one.

Qwen3.8 Flash Next

Fact Value Source
Main model parameters 125B Qwen, Qwen3.8-Flash-Next: A New Architecture
N-gram embedding tables, on top of the main model 51B the same announcement
Activated per token 6B the same announcement
Layers, hidden dimension 48 layers, hidden 2560 Qwen/Qwen3.8-Flash-Next
A published four bit MLX conversion, whole repository 104GB across eleven shards pipenetwork/Qwen3.8-Flash-Next-MLX-4bit

The 62GB in beat 035 is arithmetic, and the narration says so: “four bit arithmetic puts the main model alone around sixty two gigabytes”. 125 billion parameters at four bits is 62.5GB. It is the MAIN MODEL on its own, before the embedding tables and before any working state, and the shot labels it the main model, on its own for that reason.

The published four bit conversion of the whole repository is 104GB, which is the figure a viewer would actually meet on disk. It is not in the narration, so it is not asserted as a headline, but it is what beat 034’s embedding tables sitting on top of the main bar are drawing, and the shot carries a grey caveat rather than implying 62GB is the download.

Not put on screen