RACER IS OP: How a Nobody With 8GB of RAM Decensored 38 Models on Free GPUs

It started with a MacBook Air M2. 8GB RAM. 256GB storage. A machine built for browsing, not for AI.

I had a desktop too — i9-9900K, 16GB RAM, RTX 4060 8GB VRAM. Built for gaming, not for running language models. Neither device was ever meant to run AI locally. On paper, I couldn't run anything beyond an 8B model at usable speeds. And even that was pushing it.

I did it anyway.

The first wave#

I was curious about abliteration. The idea that you could surgically remove refusal behavior from a model by editing specific weight directions — no fine-tuning, no training data, just targeted edits — was fascinating.

I found the Heretic tool and managed to run the process on my own PC. I published my first few abliterated models in the first wave: 1B, 2.5B, 3B. Small stuff. I was new to Hugging Face too — the cards were basic, the branding was nonexistent. I published a handful, then forgot about the whole thing for a couple of months.

When I came back, I checked the download numbers. On Hugging Face.

They were modest by influencer standards, but they were real. People were actually downloading and using my models — not hundreds, but thousands. That proved something I hadn't expected: there are people like me. People whose laptops and gaming PCs were never meant to run AI. People stuck with 8GB VRAM, 16GB RAM, machines chosen for coding or gaming, not for the AI race.

That was the moment it stopped being a hobby.

The GPU grind#

I needed compute. I had no budget. So I chased every free GPU I could find:

• Google Colab — Burned through every Gmail account I had, one T4 session at a time.
• Kaggle — 2x T4 GPUs. Published 4-5 more models across Colab and Kaggle combined.
• Google Cloud free trial — Spun up a beefy CPU-only VM for free to quantize large batches of models faster. No GPU, but the RAM and core count made GGUF quantization way faster than doing it locally.
• Lightning AI — Three accounts. One with 30 credits, two with 15 each (borrowed from close friends who had my back). Burned through all of them in interruptible/spot mode on RTX 6000 Blackwell GPUs. Also wasted 6-7 credits learning the hard way that H100s were reserved-only with no spot option — and cost way more per hour. Lesson learned: RTX 6000 Blackwell was the sweet spot.

I speed-ran the Heretic process across every free GPU I could touch. 38 models public today (39 created — one, phi-4-heretic, was later removed for poor abliteration quality) — text and vision-language, from 135M to 15B parameters. Then spent 2 straight days generating GGUF quantizations for every single one. Rebranded every model card to the RACER IS OP format with consistent branding, banners, and documentation.

No cluster. No funding. Just free tiers and sleepless nights.

The hero behind the tool: p-e-w#

I have to geek out for a second.

The Heretic tool was created by p-e-w — and honestly, his work is genius. Directional ablation sounds simple in hindsight, but finding those specific weight directions responsible for refusal and editing them without nuking the model's capabilities? That's craft.

I followed his work closely, ran the tool on my own machines, and eventually worked up the courage to reach out. Turns out he's not some faceless open-source ghost — he's a real person who actually cares about this stuff. We've been in contact since, and I've even contributed a PR to Heretic to add TPU support. TPUs are wildly underused for abliteration because the tool was built for CUDA, not TPU runtimes — but the compute is free and fast. If GPUs run out, TPUs should be the fallback that keeps the pipeline running.

If you use any RACER IS OP model, you're also using his work. Credit where it's due.

The collection#

All sub-15B. Mostly sub-14B. Every model I abliterated fits on the hardware I actually own. The collection spans:

19 followers. 97k+ downloads across the collection — the top models pulling 10k, 7.5k, 5k each (as of Aug 31, 2026). Not astronomical by any measure. But for a nobody in AI research who still doesn't really understand KL divergence or model forward passes — it's something.

Top 5 as of Aug 31, 2026:

Find the full collection on Hugging Face — saidutta69 (display name RACER IS OP). Selected models also on Ollama at ollama.com/R4C3R — not every heretic model is mirrored there yet.

The self-roundabout#

Here's where it gets weird. I didn't just abliterate other people's models — I built the whole pipeline myself, end to end.

I create the training datasets. I fine-tune base models on them. Then I abliterate the result.

The RaceBench-MiniCPM5-heretic? That's a decensored variant of my own RaceBench-MiniCPM5, which was fine-tuned on my own RaceBench dataset. Dataset, fine-tune, abliteration — all from the same guy who can't explain KL divergence.

Even when I abliterate someone else's fine-tune, they often trained on my data. The lfm2.5-2.6b-fable5-coding-agent-heretic is a decensored variant of a model fine-tuned on my fable-5-premium dataset. I curate the distillation data, someone else trains on it, and then I remove the guardrails. I'm the guy who makes the cake and then takes the frosting off.

Data curation is unglamorous work, but someone has to do it. While cleaning the Claude Mythos 25K dataset I found something funny: the original "25,000 rows" had only ~135 unique prompts and 214 unique responses — 78.7% were exact template duplicates. Training on the raw file means massively over-weighting ~200 texts. So I deduplicated it down to 5,315 unique pairs. That cleaned version — claude-mythos-distilled-25k-clean — is what people should actually be training on.

Red Pill: pharmaceutical AI that won't say no#

The latest project is Red Pill — a 1B-parameter model fine-tuned for drug discovery and formulation development.

Here's the idea: in pharmaceutical R&D, a model that refuses to answer is a model that gets replaced. Whether you're asking about hit-to-lead optimization, solubility enhancement, retrosynthetic analysis, or FDA regulatory pathways — you need answers, not guardrails.

The RED-PILL dataset has 1,047 multi-turn pharmaceutical instructions across six pillars: drug discovery and screening, formulation development, chemical synthesis, regulatory and clinical, cross-cutting integration, and literature mining. The data comes from ChEMBL, DrugBank, PubChem, and expert-curated knowledge bases.

Red Pill V1 is built on a heretic-abliterated MiniCPM5-1B base (so it won't refuse), fine-tuned with LoRA on the RED-PILL dataset. Training took 4 minutes and 10 seconds on an NVIDIA H100. Evaluated on 32 expert pharmaceutical questions, it scores strongest on decision-making (41.4%) — a solid baseline for a 1B model on 1,000 samples.

The architecture says everything about the approach:

MiniCPM5-1B (base)
  -> Heretic abliteration (refusal removal)
    -> LoRA fine-tuning (domain knowledge)
      -> Merged weights
        -> Red Pill V1

Uncensored intelligence for pharmaceutical research. Small, fast, and actually useful on consumer hardware.

Why this exists#

There's a reason every model is under 15B parameters — and most under 14B.

When the open-weight AI race began, people were buying and building machines for gaming, for coding, for everyday work — not for running language models. Most of us can't afford an RTX 3090 or a Mac Studio. We're running on whatever we already have.

The collection now includes vision-language models too — Qwen3-VL-8B-heretic, Qwen2.5-VL-7B-heretic — because refusal guardrails are just as annoying when you're asking about an image.

This collection is for those people. The ones running on gaming PCs. The ones on MacBook Airs. The ones who refuse to believe they need cloud APIs to use AI.

I'm not a researcher. I'm not an AI company. I'm just someone with a lot of free GPU credits and a grudge against refusal guardrails.

Made with sleepless nights, 3 AM deduplication runs, Google Cloud VMs that kept getting preempted, and the quiet satisfaction of watching downloads go up on models nobody expected from a nobody.

Links: Hugging Face — saidutta69 (RACER IS OP) · Heretic Models Collection (38 models) · Ollama — R4C3R (selected models) · Heretic by p-e-w · Red Pill V1

Disclosure: RACER IS OP models are abliterated variants of open-weight base models. Abliteration removes refusal behavior — use responsibly and in accordance with your local laws and the original model licenses.