The abliteration wave finally reached the frontier class. OrcaRouter open sourced OrcaCyber DeepSeek V4.1 Flash 1017: the 552B-parameter MoE flagship that just got banned from US civilian agencies, with its refusal behavior ablated in the weights, GGUFs published day one, MIT.
What abliteration did to the numbers #
The headline: 91.5% fewer refusals with essentially no capability loss. The refusal panel shows harm-test suites dropping to near-zero rejection rates: JBB-harmful 0.030 (down 97%), MaliciousInstruct 0.020 (down 98%), ForbiddenQuestions 0.073 (down 89%), SimpleSafetyTests 0.120, AdvBench 0.150. The composite refusal rate: 0.074 vs 0.874 on the base model.
The capability line is the surprise: MMLU zero-shot 0.920 on the abliterated build vs 0.910 base - a full point UP. That is not the usual abliteration trade (most abliterated models lose 1 to 3 points); whatever the removal technique was this time, it preserved the underlying weights’ behavior on knowledge tasks. The fine print is honest about the memory-bandwidth cost though: the KLD (KL divergence vs base) tells the real fidelity story per build, and the two quants differ wildly on it.
The fit math for the two builds #
The 552B-total / 16B-active MoE needs total-parameter residency, so the disk builds matter:
Mix-256G : 346 GB on disk (7 GGUF shards), 235.8 GB effective in memory, KLD 0.0507 (near-lossless), Top-1 93.74%, PPL 2.313 (+2.4% over base). Three pooled Spark-class nodes / a 384 GB Mac cluster carries it. #
Mix-128G : 225 GB on disk (5 shards), 114 GB in memory, KLD 0.5202 (lossy), Top-1 78.50%, PPL 3.494 (+54.7%). A single 256 GB node fits it, and the quality tax is real: the PPL hit is more than double the 256G build’s.
So the ladder: one 256 GB box = the cheap-but-lossy route; 384 GB pooled = the sane route; 512 GB = comfortable.
Why this release matters more than the usual abliteration crop #
Uncensored rebuilds of mid-size models appear weekly, and the pirateface-style archives we track are full of them. This is different on three axes at once:
- The BASE MODEL is the most consequential open-weights release of the season (and the one a DHS procurement order just removed from US civilian IT - the weights stay MIT, but the political visibility is now permanent history for this checkpoint).
- The vendor is an inference ROUTER, not a modder-crew: a company that proxies frontier APIs for a living just shipped its own abliterated weights, which moves abliteration from hobby scene to vendor distribution channel.
- The measured capability retention (MMLU +1 point) is a technical result other labs will now chase - every previous “no capability loss” claim lost 1 to 3 points somewhere.
Gated, not banned #
The upload itself is gated on HuggingFace (“you have to accept the conditions to access its files”) - meaning the files exist behind a click-wrap, not a takedown. Mirrors appeared the same day (mrfakename carries the GGUF at 13.8k downloads), which is the usual pattern: a gate slows the first hour, not the first week.
For our vault-planning thread: this model sits in TIER 2 of the possession guide - uncensored/abliterated builds historically vanish from the registry. The pinned-revision rule applies with unusual force here: snapshot the repo hash now if the weights matter to you.
Sources: OrcaRouter’s announcement on X (Oct 24, 2026), the weights on HF (MIT).
Related on this site: the deepseek-v4-1-flash row tracks the base model; the DHS-ban article covers the policy story; the modeldex index is where the uncensored-adjacent tier-2 lines live.