Every figure, benchmark, quantisation size and architecture claim the video puts on screen, chased to a primary source. Checked September 2026.
Abliteration edits a refusal direction out of a model’s weights rather than steering the model at inference time. Maxime Labonne’s write up describes computing a refusal direction from hidden states and orthogonalising it out of the residual stream, which changes the released file rather than the prompt.
Meta’s own model card, benchmark tables:
| Benchmark | Reported |
|---|---|
| MMLU (0-shot, CoT) | 86.0 |
| IFEval | 92.1 |
| HumanEval (pass@1) | 88.4 |
| MATH (0-shot, CoT) | 77.0 |
Abliterated GGUF sizes, bartowski’s quantisations of huihui-ai’s abliterated build:
| Quant | Size |
|---|---|
| Q4_K_M | 42.52 GB |
| Q5_K_M | 49.95 GB |
| Q6_K | 57.89 GB |
mradermacher’s separate conversion of the same model lists 42.6 / 50.0 / 58.0 GB, so the figures are stable to within about a tenth of a gigabyte across repackagers.
Qwen2.5 Technical Report, Table 7, instruction tuned results for Qwen2.5-32B-Instruct:
| Benchmark | Reported |
|---|---|
| MMLU-Pro | 69.0 |
| MATH | 83.1 |
| HumanEval | 88.4 |
| LiveCodeBench | 51.2 |
| Arena-Hard | 74.5 |
Abliterated GGUF sizes, tensorblock’s quantisations of huihui-ai’s abliterated build:
| Quant | Size |
|---|---|
| Q4_K_S | 17.494 GB |
| Q4_K_M | 18.488 GB |
| Q5_K_S | 21.084 GB |
| Q5_K_M | 21.665 GB |
Repackagers disagree more here than they do on Llama. mradermacher’s conversion of the same abliterated model lists Q4_K_M at 20.0 GB and Q5_K_M at 23.4 GB. The video uses the tensorblock figures, which are the smaller of the two sets.
Mistral’s own model card, 3.2 against 3.1:
| Benchmark | Small 3.1 | Small 3.2 |
|---|---|---|
| Arena Hard v2 | 19.56% | 43.1% |
| HumanEval Plus (pass@5) | 88.99% | 92.90% |
| Internal instruction following accuracy | 82.75% | 84.78% |
The card also states Small 3.2 “is better at following precise instructions” and that its “function calling template is more robust”.
Abliterated Q4_K_M GGUF is about 14.3 to 14.4 GB depending on the repackager. mradermacher’s conversion of huihui-ai’s abliterated build lists Q4_K_M at 14.4 GB, with 16.9 GB for Q5_K_M and 19.4 GB for Q6_K.
Google’s own model card:
| Benchmark | Reported |
|---|---|
| MMLU (5-shot) | 78.6 |
| MBPP (3-shot) | 65.6 |
| MMMU (pt) | 56.1 |
Context window 128K tokens for the 4B, 12B and 27B sizes. Multimodal, accepting image and text input. Multilingual support in over 140 languages.
Abliteration. Maxime Labonne’s model card states Gemma 3 was much more resilient to abliteration than Qwen 2.5, and that the release computes a refusal direction per layer independently, with a refusal weight of 1.5, rather than applying one direction globally.
Abliterated GGUF sizes, published in the same author’s GGUF repo:
| Quant | Size |
|---|---|
| q3_k_m | 13.4 GB |
| q4_k_m | 16.5 GB |
| q5_k_m | 19.3 GB |
| q6_k | 22.2 GB |
Qwen’s own model card:
| Benchmark | Reported |
|---|---|
| Terminal Bench 2.1 | 73.0 |
| SWE-bench Pro | 61.7 |
| QwenSWEBench | 79.0 |
| LiveCodeBench v6 | 90.3 |
Context window 262,144 tokens natively, extensible to 1,000,000. A native vision language model that understands images and video. Thinking mode on by default and disableable per request, with reasoning depth tunable. Hybrid architecture of Gated DeltaNet linear attention layers interleaved with full attention layers.
Two figures the narration states could not be confirmed, so neither appears on screen.
The abliterated MTPLX quantisation sizes. The script gives 30.4 GB at 8 bit, 23.7 GB at 6 bit and 16.9 GB at 4 bit. No published abliterated build lists those. MTPLX is a real project, a native MTP speculative decoding runtime for MLX on Apple Silicon, but it publishes no abliterated build: its Qwen 3.8 27B releases are Optimized Speed, Bare Speed and Optimized Quality, none of them abliterated. The nearest real abliterated MLX build is PocketAiHub’s, which lists 27.50 GB at 8 bit, 21.24 GB at 6 bit and 14.98 GB at 4 bit.
The M3 Ultra Mac Studio speed. The script gives roughly 46 tokens per second for the 6 bit build on a 24,000 token agentic prompt, and over 50 for the 4 bit build. No source reports either. PocketAiHub’s abliterated MLX card measures on an Apple M5 Max with 128 GB of unified memory at a 4,105 token prompt, giving 33.2 tokens per second at 4 bit and 24.7 at 6 bit, and calls these single local runs rather than guarantees. MTPLX reports relative speedups, 1.6x on a 16 GB M4 Mac mini and 2.24x on an M5 Max, and no absolute figure on an M3 Ultra.
The claim that a 27B class model runs at usable interactive speed on Apple Silicon is well supported. The specific numbers are not, so the shots covering those two lines draw the mechanism and print no figure.