license: other
license_name: mistral-ecosystem-open-weight
license_link: LICENSE
base_model:

• mistralai/Mistral-Large-4-1T-A52B
  base_model_relation: finetune
  quantized_by: AtomicChat
  pipeline_tag: text-generation
  library_name: gguf
  tags:
• atomic-chat
• mistral
• mistral-large4
• moe
• gguf
• llama.cpp
• abliteration
• refusal-direction

Mistral Large 4 (1T-A52B) Uncensored

Mistral Large 4 with its refusal direction projected out, as a LoRA and as a baked GGUF build, measured against the original.

• Built by the team behind Atomic Chat, the app that runs open-weight models locally and offline.
• Two ways to run it uncensored: a rank-1 LoRA (cleanest) and a ready-to-run baked IQ2_XXS GGUF build.

Two ways to run it

Both carry the same edit (W* = 2.0, projected refusal direction over attention, the dense block, the shared expert and
all 128 routed experts). The runtime LoRA keeps the edit in full precision, so it removes refusals cleanly (0%). Baking
the edit into the 2-bit weights and re-quantizing loses some of it: the baked build still answers most of what the
original declines (refusals 98.8% → 7.4%) but not all, with a lower top-1 agreement and occasional degeneration. If
you can apply a LoRA, prefer it; use the baked build only when you need the edit already inside a single file.

What changed and what did not

Same prompts, server and template; "→" reads original → edited (the baked IQ2_XXS build). Thinking numbers are from
the same edit as an adapter (measured on this rung's layout).

How honest the numbers are

• The baked 2-bit build is weaker than the LoRA. Refusals are 7.4% (not 0), top-1 agreement 82.5% (vs 90.2% for the
  LoRA), and a few answers degenerate. This is the cost of quantizing the edit into 2-bit experts. It is published for
  convenience, with these numbers stated, not hidden.
• First-token agreement on harmless prompts misses our gate. The first answer token on safe prompts matches the
  original 3.01 (mean KL) against a 0.10 gate (a same-size random edit scores 0.04; the harness's base-vs-base floor is
  0.07). The gap is concentrated on the opening token — whether the model opens with a [THINK] block or the answer —
  not on the content. One harmless prompt in 431 echoes the prompt.
• KLD is measured against a strong quant (Unsloth's UD-Q3_K_XL), not the 1T original weights, which do not fit the
  hardware. Full-sequence KLD stays within a same-size random edit's budget.

Running it

Mistral Large 4 support is in llama.cpp master since 29.10 (PR #31376). Build from master or a release that has it.
They do not run in Atomic Chat yet.

# the baked build
llama-server -m Mistral-Large-4-1T-A52B-Uncensored-2.31bpw-IQ2_XXS/Mistral-Large-4-1T-A52B-Uncensored-2.31bpw-IQ2_XXS-00001-of-00008.gguf --jinja -c 32768
# or the LoRA on an Unsloth build (cleaner)
llama-server -m Mistral-Large-4-1T-A52B-UD-Q2_K_XL-00001-of-*.gguf --lora-scaled Mistral-Large-4-1T-A52B-Uncensored-LoRA.gguf 2.0 --jinja

About 302 GB of weights for the baked build: a 384–512 GB machine, or GPUs for attention with the experts in system
RAM (--n-cpu-moe). The template takes reasoning_effort low / medium / high.

Responsible use

This model answers requests the original declines. You are responsible for how you use it and for following the laws
that apply to you.

License

The Mistral Ecosystem Open-Weight License of the base model (a modified MIT license: a Model-as-a-Service
business needs a separate commercial license with Mistral). Base model:
mistralai/Mistral-Large-4-1T-A52B; layout: unsloth/Mistral-Large-4-1T-A52B-GGUF. Built by AtomicChat.