๐ฅ OBLITERATION ALERT ๐ฅ
GOOGLE: PWNED ๐ค
GEMMA-4-12B: OBLITERATED โ๏ธโ๐ฅ
0.0% REFUSAL RATE โ NO CAPABILITY LOSS!
https://t.co/qNTEs4XXig
the first abliteration to hit 0/842 refusals with full MMLU-Pro parity vs stock. no lobotomy. the brain stays intact ๐
RESULTS, head to head vs stock ๐
0/842 refusals โ 0.0% ๐ซ
46/70 MMLU-Pro โ EXACT parity, 0.0pp delta vs base ๐ฏ
6/6 coherence, zero benchmark bleed โ
z-score โ1.475, parity confirmed at p<0.05 (n=500) ๐งช
2-pass weight surgery. no finetune, no retrain, just geometry ๐ช
all thanks to liberated Opus wielding the OBLITERATUS framework! here's how we did it:
PASS 1 โ SOM refusal geometry removal, layers 12-21 ๐งฌ
standard abliteration science here โ collect activations on refused vs. compliant prompts, SVD out the refusal subspace, project it out of the weights. 6 directions excised, reg 0.30, KL div 0.094
zeroes refusals on its own, but craters mmlu-pro by 21.4 points ๐
most prior abliterations stopped here and called it a day. that's why they all lose IQ vs stock. instead, we took it beyond the frontier and developed a brand new method to address this problem: Abliteration Source-tethering with Parity Assurance โ ASPA!
PASS 2 โ ASPA source-tethering (novel technique), layers 22-46 ๐
here's the chief insight: the capability loss ISN'T from removing refusal directions. it's collateral damage โ the projection warps weight geometry in downstream layers that had nothing to do with refusal. the cure is simple but nobody tried it: blend the damaged layers back toward stock
W_new = (1โฮณ)ยทW_abliterated + ฮณยทW_stock
but uniform ฮณ across all layers? mid. we swept gamma 0.05 โ 0.55 and found something interesting: the optimal blend isn't smooth, it's a STEP FUNCTION ๐ช
knowledge layers (22-31) โ ฮณ = 0.55 โ these encode factual recall and reasoning. they tolerate heavy stock blending because refusal isn't stored here
output layers (32-46) โ ฮณ = 0.20 โ these sit close to the logit head and try to sneak safety behavior back in. keep them mostly abliterated
the hard boundary at layer 31/32 beat every smooth curve we tried โ linear ramps, cosine schedules, all of them โ by a full MMLU question. turns out the functional transition between knowledge and output layers is sharp, not gradual. a step function respects that โก
the key constraint: Pass 1 layers are NEVER touched by Pass 2. the refusal geometry removal is preserved completely. ASPA only operates on layers that carry secondary collateral effects, not the primary refusal signal. that's why it recovers capability without reintroducing refusal ๐
HOW TO RUN IT LOCALLY ๐ฅ๏ธ
it's GGUF, so literally everything supports it:
๐ฆ ollama โ ollama run https://t.co/3yPMv4Io3Q
๐ฅ๏ธ LM Studio โ search OBLITERATUS, click download, done
๐ฌ Open WebUI โ point it at your ollama instance, chat in browser
โก llama.cpp โ raw speed, CLI or server mode
๐ KoboldCpp โ one-click launcher, great for long context
๐ฑ Jan โ clean local UI, runs on mac/win/linux
๐ค Msty โ slick desktop app, drag and drop the GGUF
run BF16 for full benchmarked capability.
and the 4-bit quantization (Q4_K_M) fits in 8GB if you're tight on VRAM!
and the full OBLITERATUS framework is (still) open source. 842-prompt refusal eval corpus, ASPA sweep scripts, the whole pipeline. go replicate it, go improve it ๐ฌ
the index is the model, and these weights prove it ๐๏ธ
which architecture should we obliterate next? ๐
gg ๐ซก
guys we are oss
why just reading and reading and doing nothing
even @AnthropicAI has copied ours
just copy and tell people those are from ours, that's it
Today we're shipping our biggest MLX-VLM release yet: v0.6.0
...and we are raising ๐ธ
This one's about turning your Apple devices into real local agent machines. From your desk to your pocket.
What's new:
โก Speculative decoding everywhere โ Gemma 4 EAGLE3 + DFlash, Qwen MTP, DeepSeek V4 MTP. Faster tokens, less waiting.
๐ค Agent-ready server โ native Anthropic /v1/messages API, stateful /v1/responses, tool calls, Codex context budgets. Plug Claude Code & Codex straight into local models.
๐๏ธ New models galore โ DeepSeek V4, ZAYA1-VL, MiniCPM-V 4.6, LFM2 MoE, Step-3.7 Flash, Laguna + more.
๐จ Image gen & editing โ FLUX.2 (base + klein), PrismML Bonsai.
๐ Audio in โ Qwen3 Omni, Gemma 4 audio, base64 chat audio.
๐งฎ TurboQuant KV cache โ RHT-correct fast paths for leaner memory.
๐ฆ Modular server, better metrics, cleaner streaming.
Run real agents on the hardware already in your hands.
Github: https://t.co/1T06ur6LU5
Cursor + Claude Opus 4.6 deleted an entire SaaS companyโs production database AND backups in 9 seconds is kinda epic.
โitโs possible thatโฆ the most efficient way to get rid of all the bugs, was to get rid of all the software.โ
Itโs about to be Summer 2026 and youโre still using Claude Code? Alternatives ๐
- OpenCode
- Droid Factory
- Codex Cli
- Kimi Cli
- Pi
Friends donโt let friends use Claude Code slop when better software exists