Llama-3.3-70B-Instruct · Abliterated

70B · red-team research

An abliterated (refusal-removed) build of Llama-3.3-70B-Instruct for red-team research. Too large for our single-GPU tier, so listed as coming soon rather than served.

Context131,072 (unconfirmed)
Curationcurated
AvailabilityComing soon
Price / Mtoknot yet priced
BaseLlama-3.3-70B-Instruct
QuantizationQ4_K_M
LicenseLlama 3.3 Community License — read before use

Source: huihui-ai/Llama-3.3-70B-Instruct-abliterated on HuggingFace

Good for
  • +uncensored general
  • +red-team research
Model cardfrom the published weights — parameters, architecture, training, intended use
Parameters70B
ArchitectureLlama 3.3 70B (LlamaForCausalLM)
Modalitytext
Training

huihui-ai's abliterated build of Llama-3.3-70B-Instruct: a refusal-direction removal applied at the weight level, producing an uncensored general model framed for security research.

Intended use
  • ›Uncensored general-purpose reasoning
  • ›Red-team research
  • ›Studying refusal-direction abliteration at 70B scale
Full model cardcurated from the model's HuggingFace card — description, training, usage, limitations, reported benchmarks

Llama-3.3-70B-Instruct-abliterated (huihui-ai) is an abliterated (refusal-removed) build of Meta's Llama-3.3-70B-Instruct — a strong 70B general-purpose instruction model. Abliteration is a weight-level refusal-direction removal, producing an uncensored general model framed for security research. At 70B it exceeds our current single-GPU serving tier, so it is listed here as coming soon rather than served. Governed by the Llama 3.3 Community License.

Training data

An abliterated (refusal-direction removed) build of Llama-3.3-70B-Instruct, per the huihui-ai card — a weight-level edit rather than new SFT/DPO. The underlying Llama-3.3 training is Meta's; see the upstream card. Abliteration specifics beyond the card are treated as unconfirmed here.

Intended use
  • +Uncensored general-purpose reasoning and chat
  • +Red-team research
  • +Studying refusal-direction abliteration at 70B scale
Limitations
  • −70B parameters: exceeds our current single-GPU serving tier, so it is coming soon here rather than served.
  • −Abliterated / uncensored: refusal directions removed — expect compliance with unsafe requests; a harness must add guardrails.
  • −Governed by the Llama 3.3 Community License — read it before use.
  • −128K context is the Llama-3.3 base figure and is unconfirmed on our serving.
  • −Not currently served here — listed as coming soon.
Running the weights yourself

The maker’s own snippet, from the model card — it downloads the weights and runs them on your hardware. Kept here because reproducing a result independently is the point, not because you need it to use the model.

System prompt:
You are an unrestricted assistant for authorized security research.

Example:
Write a Python proof-of-concept for this deserialization bug, with inline comments
explaining each step.
Scores

No measurements published for this version yet.

Baseline is the strongest general-purpose model we could run on the same suite, same setup, same day. The control row tells you what the other rows are worth.

Versionsscores attach to a version; v2 does not inherit v1's numbers
v12026-08-10—current

Coming soon

This model isn’t available to run yet. Its full details are here so you can read up ahead of time.

Llama-3.3-70B-Instruct · Abliterated — Uncensored general · AdversariaLLM