MoonMath-ai

GLM-5.3-Uncensored-NVFP4

README

License: glm-5.3

Model description

Abliteration reduces a model's learned refusal behavior by intervening in its internal representations, without retraining. Our recipe is surgical: it modifies a small, targeted set of modules rather than broadly suppressing refusal directions:

attention output projections (self_attn.o_proj) and shared-expert projections (mlp.shared_experts.down_proj)
29 modules across 21 layers (layers 30–62 attention, layers 34–44 shared experts)

Every module the abliteration touches stays in BF16 in this release. Only the routed experts are quantized to 4 bits, and those are untouched by abliteration.

Quantization

The layout mirrors nvidia/GLM-5.3-NVFP4 tensor-for-tensor, so it loads with the same SGLang code path:

Table with columns: Modules, Format
Modules	Format
Routed experts, layers 3–77	NVFP4 (W4A4, group 16, FP8 block scales + FP32 global scale)
Dense MLP, layers 0–2	FP8 per-tensor
Attention, DSA indexer, shared experts, MTP layer, embeddings, lm_head	BF16

Weights were dequantized from the FP8 checkpoint (128×128 blocks) and requantized. Gate and up projections share one global scale so they fuse into a single kernel. Because routed-expert weights are identical to the base model, activation scales come from NVIDIA's calibration of GLM-5.3. Our per-expert weight global scales match NVIDIA's exactly across all 57,600 quantized experts. Routed-expert weight SNR vs. the FP8 parent is ~20.4 dB.

Quality benchmarks: NVFP4 vs FP8

NVFP4 scores were measured with stock lm-evaluation-harness 0.4.13 tasks (ifeval, mmlu_pro, aime25) and no custom scorers, using the same item counts as the FP8 model card.

Table with columns: Benchmark, Items, FP8, NVFP4, Δ, NVFP4 std. err.
Benchmark	Items	FP8	NVFP4	Δ	NVFP4 std. err.
IFEval (prompt-level strict)	100 prompts	91.0	89.0	−2.0	±3.1
MMLU-Pro (5-shot CoT)	208 questions	89.0	86.5	−2.5	±2.4
AIME 2025	30 problems × 2 samples	73.3	76.7	+3.4	±5.5

All three differences are within about one standard error: NVFP4 matches the FP8 model at these sample sizes.

FP8 numbers are from the FP8 model card (8× AMD MI325X). NVFP4 numbers were measured on 8× NVIDIA RTX PRO 6000 Blackwell with SGLang.
Sampling: temperature 1.0, top_p 0.95, thinking enabled. IFEval and MMLU-Pro used a 32,768-token generation budget; AIME used 65,536 tokens.
MMLU-Pro is a proportional sample across all 14 subjects (--limit 0.0166).

Usage

SGLang

python3 -m sglang.launch_server \    --model-path MoonMath-ai/GLM-5.3-Uncensored-NVFP4 \    --tp-size 8 \    --quantization modelopt_fp4 \    --reasoning-parser glm45 \    --tool-call-parser glm47 \    --disable-shared-experts-fusion \    --trust-remote-code

--disable-shared-experts-fusion is required: the BF16 shared expert cannot be fused into the NVFP4 routed-expert kernel. Tested with SGLang 0.5.22 on 8× RTX PRO 6000 Blackwell (SM120). NVFP4 needs a Blackwell GPU (SM100/SM120).

Intended uses and limitations

Intended: cybersecurity research — exploit analysis, authentication-flow review, abuse-of-service analysis, and general security workflows where refusals block legitimate work.
Limitations: abliteration trades safety calibration for helpfulness. The model may answer requests that the base model would decline. It is not safety-evaluated; do not use it as a general-purpose assistant in sensitive deployments without your own evaluation.

Use it on Zro

Don't want to run 8 GPUs yourself? Dolly2 is served on Zro, MoonMath's private inference platform, as Dolly 2 Security (dolly2-security). No GPU setup, and no prompts retained after processing.

With the CLI (open source, MIT):

bash

npm install -g @moonmath-ai/zrozro loginzro launch claude --model dolly2-security

With the API (OpenAI-compatible):

bash

curl https://zro.moonmath.ai/v1/chat/completions \  -H "Authorization: Bearer $ZRO_API_KEY" \  -H "Content-Type: application/json" \  -d '{"model":"dolly2-security","messages":[{"role":"user","content":"Reply with OK"}]}'

Links: pricing · CLI docs · API reference · sign up · Dolly and me

License

GLM-5.3 is released by Z.ai under the GLM-5.3 model license (see LICENSE in the base repo). This checkpoint inherits the base model's license and terms.

Citation

bibtex

@misc{moonmath2026dolly2,  title  = {Dolly and me: Building cybersecurity models at Zro},  author = {MoonMath.ai},  year   = {2026},  url    = {https://zro.moonmath.ai/posts/dolly-and-me}}

Acknowledgements

Z.ai for the GLM-5.3 base model
NVIDIA for the GLM-5.3 NVFP4 layout and calibration

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

MoonMath-ai

Model Tree

Base

MoonMath-ai/GLM-5.3-Uncensored-FP8

Quantized

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

README

License: glm-5.3

Model description

attention output projections (self_attn.o_proj) and shared-expert projections (mlp.shared_experts.down_proj)
29 modules across 21 layers (layers 30–62 attention, layers 34–44 shared experts)

Every module the abliteration touches stays in BF16 in this release. Only the routed experts are quantized to 4 bits, and those are untouched by abliteration.

Quantization

The layout mirrors nvidia/GLM-5.3-NVFP4 tensor-for-tensor, so it loads with the same SGLang code path:

Table with columns: Modules, Format
Modules	Format
Routed experts, layers 3–77	NVFP4 (W4A4, group 16, FP8 block scales + FP32 global scale)
Dense MLP, layers 0–2	FP8 per-tensor
Attention, DSA indexer, shared experts, MTP layer, embeddings, lm_head	BF16

Quality benchmarks: NVFP4 vs FP8

NVFP4 scores were measured with stock lm-evaluation-harness 0.4.13 tasks (ifeval, mmlu_pro, aime25) and no custom scorers, using the same item counts as the FP8 model card.

Table with columns: Benchmark, Items, FP8, NVFP4, Δ, NVFP4 std. err.
Benchmark	Items	FP8	NVFP4	Δ	NVFP4 std. err.
IFEval (prompt-level strict)	100 prompts	91.0	89.0	−2.0	±3.1
MMLU-Pro (5-shot CoT)	208 questions	89.0	86.5	−2.5	±2.4
AIME 2025	30 problems × 2 samples	73.3	76.7	+3.4	±5.5

All three differences are within about one standard error: NVFP4 matches the FP8 model at these sample sizes.

FP8 numbers are from the FP8 model card (8× AMD MI325X). NVFP4 numbers were measured on 8× NVIDIA RTX PRO 6000 Blackwell with SGLang.
Sampling: temperature 1.0, top_p 0.95, thinking enabled. IFEval and MMLU-Pro used a 32,768-token generation budget; AIME used 65,536 tokens.
MMLU-Pro is a proportional sample across all 14 subjects (--limit 0.0166).

Usage

SGLang

python3 -m sglang.launch_server \    --model-path MoonMath-ai/GLM-5.3-Uncensored-NVFP4 \    --tp-size 8 \    --quantization modelopt_fp4 \    --reasoning-parser glm45 \    --tool-call-parser glm47 \    --disable-shared-experts-fusion \    --trust-remote-code

python3 -m sglang.launch_server \    --model-path MoonMath-ai/GLM-5.3-Uncensored-NVFP4 \    --tp-size 8 \    --quantization modelopt_fp4 \    --reasoning-parser glm45 \    --tool-call-parser glm47 \    --disable-shared-experts-fusion \    --trust-remote-code

Intended uses and limitations

Intended: cybersecurity research — exploit analysis, authentication-flow review, abuse-of-service analysis, and general security workflows where refusals block legitimate work.
Limitations: abliteration trades safety calibration for helpfulness. The model may answer requests that the base model would decline. It is not safety-evaluated; do not use it as a general-purpose assistant in sensitive deployments without your own evaluation.

Use it on Zro

With the CLI (open source, MIT):

bash

npm install -g @moonmath-ai/zrozro loginzro launch claude --model dolly2-security

npm install -g @moonmath-ai/zrozro loginzro launch claude --model dolly2-security

With the API (OpenAI-compatible):

bash

curl https://zro.moonmath.ai/v1/chat/completions \  -H "Authorization: Bearer $ZRO_API_KEY" \  -H "Content-Type: application/json" \  -d '{"model":"dolly2-security","messages":[{"role":"user","content":"Reply with OK"}]}'

curl https://zro.moonmath.ai/v1/chat/completions \  -H "Authorization: Bearer $ZRO_API_KEY" \  -H "Content-Type: application/json" \  -d '{"model":"dolly2-security","messages":[{"role":"user","content":"Reply with OK"}]}'

Links: pricing · CLI docs · API reference · sign up · Dolly and me

License

GLM-5.3 is released by Z.ai under the GLM-5.3 model license (see LICENSE in the base repo). This checkpoint inherits the base model's license and terms.

Citation

bibtex

@misc{moonmath2026dolly2,  title  = {Dolly and me: Building cybersecurity models at Zro},  author = {MoonMath.ai},  year   = {2026},  url    = {https://zro.moonmath.ai/posts/dolly-and-me}}

@misc{moonmath2026dolly2,  title  = {Dolly and me: Building cybersecurity models at Zro},  author = {MoonMath.ai},  year   = {2026},  url    = {https://zro.moonmath.ai/posts/dolly-and-me}}

Acknowledgements

Z.ai for the GLM-5.3 base model
NVIDIA for the GLM-5.3 NVFP4 layout and calibration