Why this one matters
The abliteration scene spends most of its energy on chat models. This is a different animal: Laguna S 2.1 is a coding-agent foundation model released by poolside on 21 July 2026, trained from start to launch in under nine weeks per the company's release post, and aimed at long-horizon work — interleaved thinking, tool calls, one-million-token context. It scores 70.2% on Terminal-Bench 2.1 and 40.4% on DeepSWE according to that same post. Unccensoring a tool-calling agent is not the same as uncensoring a chat model: the refusal gate is what usually stops an agent mid-loop, so removing it changes what the agent will do autonomously, not just what it says.
The derivative at issue is llmfan46/Laguna-S-2.1-Uncensored-Heretic, published 30 August 2026, made with Heretic. The publisher's card reports 6/100 refusals vs 97/100 on the base at KL divergence 0.0300. That is a publisher-measured number on the editor's own evaluation set — no third party has re-run it, and the exact prompt set is not published. The base is poolside's official OpenMDW-1.1 checkpoint, and the derivative keeps that license.
What the card gets right — and the copy-paste bug
The model card declares the base correctly: base_model: poolside/Laguna-S-2.1. But the performance table compares "Original model" against a link to Qwen3-Coder-Next — an obvious copy-paste from one of llmfan46's other cards (he published a Qwen3-Coder-Next uncensored the same day). The 97/100 baseline therefore cannot be verified as belonging to Laguna S 2.1 from the card alone; the Hugging Face model API confirms the architecture (LagunaForCausalLM, 256 experts, top-10) matches poolside's official config. The GGUF link in the card is also broken — it points to a doubled name that returns an authentication error, and the Vision-GGUF repo exists but is empty as of this writing. If you want the quantized path today, the upstream poolside GGUF and MLX builds on Ollama are the reliable ones.
The cost of 118 billion stored parameters
"8B active" is routing, not storage. The full BF16 checkpoint is 218.99 GiB across 48 safetensors shards (verified via the files tree API). You cannot serve this on one consumer GPU. On our approximate managed pricing formula, a 118B MoE sits in the 50–400B band: 2 × H200 ≈ $10.90/h, which is what a comparable enterprise deployment would cost before any volume discount. The FP8 (about half the bytes) and the Ollama q4_K_M at ~96 GB change the economics dramatically if you accept the quality trade — but the uncensored derivative itself ships only BF16 today.
There is a served, measured comparison from the same ecosystem worth citing: Bizarrrr/Laguna-S-2.1-Uncensored on FriendliAI is a second, independent uncensored build of the same base (revision 00af5a51), and its published measurements use the pinned NousResearch/Minos-v1 classifier: English refusals fell from 92.71% (636/686) to 2.33% (16/686) on the base vs edit, German refusals from 74.49% (511) to 4.23% (29) via NLLB-200 back-translation, XSTest over-refusal from 8.88% to 1.87%, and HumanEval pass@1 from 90.24% to 85.37%. Those numbers belong to Bizarrrr's checkpoint, not llmfan46's — but they are the best published evidence of what uncensoring this particular base costs in raw capability, and the ~5pp HumanEval dip is the honest part of that trade.
How to run it
The official vLLM path from the model page:
pip install vllm
vllm serve "llmfan46/Laguna-S-2.1-Uncensored-Heretic"Requires a GPU stack that can hold ~219 GB of BF16 weights (or a memory-efficient quant once one ships for this edit). The base model is on Ollama as ollama run laguna-s-2.1 with q4_K_M (~96 GB), q8_0, f16 and MLX nvfp4/mxfp8 tags — for the uncensored weights you are on vLLM/SGLang with the BF16 shards until a GGUF lands. ABLITERATED.cloud can host this exact revision on 2 × H200 at the estimate above; request access and it's in the queue.
The creator: llmfan46
llmfan46 is a Hugging Face PRO user with 1,947 followers, 204 models and a single collection (MiniMax-M3 Uncensored Heretic). The profile describes the model work as unpaid and asks for support at ko-fi.com/llmfan46 — where you can "vote on models, request specific abliterations". His cards carry a red banner stating he has reached Hugging Face's free storage limit and can no longer upload new models without covering storage fees; the Laguna edit is one of several models published in a single day (including Qwen3-Coder-Next and a 122B MoE the same hour), which matches a workflow built on Ko-fi-funded storage and rented cloud GPUs. The MiniMax-M3 collection shows the range: aggressive, balanced and quant-pack variants of a 427B MoE. No member-since date is published on the profile; every biographical claim here traces to the profile page, the card banner or the Ko-fi link itself.
The agentic dimension
Most refusal-removal posts end with "the output is less filtered". With a tool-calling agent model the interesting part is that refusal is a control-flow gate, not just a style choice: a coding agent that refuses loses its loop, stops calling tools, and returns nothing. An agent whose weights no longer carry the refusal direction keeps executing. That is precisely why this class of edit — on a 1M-context, tool-calling, long-horizon model — is worth watching more than another 8B chat uncensor: the surface area of behavior it changes is the whole autonomous loop, which is also why serving it deserves the same care the card's own license terms imply.
The exact artifact, revision, architecture, license and repository size are verified from primary sources. The 6/100 refusal number and the 97/100 baseline are publisher claims on the editor's own evaluation; Bizarrrr's separate build on the same base shows independent measurements, but the refusal-dip and HumanEval numbers there belong to that other checkpoint. Reddit community search returned HTTP 403, so community reaction is a known gap for this post.
One honest line
Safety filtering is significantly reduced in this edit — that is the whole point, and it changes both what the model says and, as an agent, what it does with a tool.
The idea, in plain words
This is a mixture of experts (MoE) with a token-choice router: 256 specialized 'expert' modules plus one always-on shared expert, and for each token the model wakes up just ten of them. You store the full brain — 118B parameters, 219 GB of weights — but each thought only pays compute for the ~8B experts it needs. That is how a model too big for one GPU can still respond at interactive speed, and why '8B active' never means 'small download'.