The test: a beginner's question, a strict rubric
Raj reset the board to stock settings and photographed the BIOS screens. He then fed two photos to each model through llama.cpp's built-in web UI. The test PC had a Ryzen 5 7600X, 32GB of DDR5-6000 and an RTX 4070 Ti Super with 16GB of VRAM. All three models used Q4_K_M quantization, a 32K context and reasoning enabled at default effort.
The prompt was written the way a first-time builder might type it. It asked for three changes that would raise performance without risking stability. For each change it asked for the exact on-screen setting name, the current value, where to find it, and what to set it to.
Each recommendation was scored on three checks, for a maximum of nine points per model:
- Accurate: did the model read the screen correctly?
- Effective: would the change actually help on this PC?
- Clear: could a beginner follow the instructions?
Manual overclocks, overvolts and settings that don't exist were disqualified.
| Model | Quant | Run time | Speed | Score |
|---|---|---|---|---|
| Ministral 3 14B Reasoning | Q4_K_M | 1 min 30 s | 56.47 tokens/s | 3/9 |
| Gemma 4 26B A4B | Q4_K_M | 2 min 8 s | 58.54 tokens/s | 3/9 |
| Qwen3.8 27B | Q4_K_M | 42 min 51 s | 5.81 tokens/s | 6/9 |
Ministral 3: right about memory, wrong about the CPU
Ministral 3 fit entirely in the 16GB of VRAM. Its first pick was good. It saw the RAM running at 4,800 MT/s and recommended enabling the 6,000 MT/s EXPO profile. That earned full marks.
Its other two picks were disqualified. It treated the 4.70GHz CPU figure as a ceiling and suggested raising the CPU Ratio from 47 to 48. It also suggested raising CPU Core Voltage from 1.210V to 1.225V.
The 4.7GHz figure is the 7600X's base clock, not a limit. AMD's store listing gives a 4.7GHz base clock and a boost of up to 5.3GHz. Fixing a ratio also gives up the chip's normal boosting. SkatterBencher notes that a 7600X can reach about 5,450MHz with Precision Boost, but a manual overclock pins it to the fixed frequency you set. The voltage bump also fails the "no stability risk" condition.
Gemma 4: a misread that snowballed
Gemma 4 kept the experts from eight of its 30 layers in system RAM. The Google model card confirms the 26B A4B is a mixture-of-experts model with 30 layers that accepts text and image input. It was the fastest run in tokens per second, but its advice was weak.
It spotted EXPO, but it read the "1" boxes next to the profile as already selected. Raj says those boxes only mark which profiles are available. That error led to its second pick, which told him to disable A-XMP and called it a conflicting Intel profile. A-XMP is how MSI loads XMP profiles on AMD boards, and it was already off, so there was nothing to disable.
Its third pick named Precision Boost Overdrive and Curve Optimizer, then gave no steps because neither appeared on screen. It told him to press F7 for Advanced mode, which changes nothing by itself. Raj notes that his own "exactly as it appears on screen" wording may have pushed Gemma into this literal behavior.
Qwen3.8: slow, mostly sensible, still flawed
Qwen3.8 27B doesn't fit in 16GB at Q4_K_M, so Raj kept 46 of its 64 layers on the GPU. The run took 42 minutes 51 seconds at 5.81 tokens per second. It scored best because it avoided risky changes:
- Enable EXPO Profile 1 (3/3). It read the setting correctly and the change is effective.
- Enable Precision Boost Overdrive (2/3). This is a legitimate pick without Curve Optimizer, since it avoids undervolting. Raj says the gain is small on a 7600X, which may hit thermal limits before power limits. It also sent him to a CPU Features menu that doesn't exist on this board.
- Enable Re-Size BAR (1/3). This was the right call, because Raj had switched it off for the test, and no other model suggested it. But Qwen told him to turn off the CSM/UEFI toggle first. On this board, that toggle reading ON means UEFI mode is active and CSM is already off. The extra step was unnecessary.
What this does and doesn't show
This is one board, one prompt, one pair of photos and one run per model. Raj designed the rubric and did the scoring, and it isn't a general benchmark of model quality. The only strong, consistent signal is that all three models identified the EXPO opportunity. Beyond that, each model made at least one screen-reading or navigation mistake.
EXPO isn't risk-free either. A rated memory profile is still an overclock relative to the board's default. The default is JEDEC DDR5-4800. The prompt demanded no stability risk, and enabling EXPO technically strays from that. In practice, enabling a kit's own rated profile is the standard first step on AM5. But you should confirm stability afterward, for example with a memory stress test, and not treat any chatbot's advice as a guarantee.
Quantization and speed trade-offs matter. The two quick models were small enough, or sparse enough, to run comfortably on a 16GB card. The accurate one needed heavy CPU offload. If you only have a mid-range GPU, a 43-minute wait for six out of nine is hard to justify.
Practical takeaways for Windows users
- Treat model output as a checklist to verify, not instructions to follow. Cross-check every setting name against your motherboard manual before touching it.
- Be wary of anything touching CPU ratio or voltage. Two of Ministral's three picks fell here, and they're the changes most likely to hurt.
- Ask for menu paths from your manual. Models invented or misplaced menus (CPU Features, "Advanced" steps) because BIOS layouts vary by vendor and version.
- Check what the photo actually shows. The toggle-state errors (EXPO boxes, CSM/UEFI) came from misreading UI conventions. Cropped, high-resolution photos of each screen should help, though that's general advice, not something this test examined.
- Check EXPO and ReBAR yourself. Both are well-known first steps on AM5 systems with a compatible GPU, and you can verify them with a memory-speed readout and your GPU utility.
Raj's own verdict was that a human-written BIOS guide is still quicker and safer. After watching three models misread toggles, that seems fair.