Jiunsong

SuperGLM-5.2-abliterated-NVFP4

Release Highlights

Table

Base architecture	GLM-5.2, 753B total / 40B activated MoE, up to 1M context
Ultimate NVFP4 parent	`nvidia/GLM-5.2-NVFP4@aec724e`
Immediate fusion parent	`Jiunsong/SuperGLM-5.2-abliterated-NVFP4@7d93e27`
This fused release	`90474e81c9351b5855d9b2385ecafc49c349a896`
Release format	Approximately 465 GB, official mixed-precision ModelOpt NVFP4 layout
OBLITERATUS	Two original + two residual directions, strength 1.5, layers 16–77
SuperTune recovery	Original rank-32 quality recovery plus incremental rank-64 teacher-forced LM-head recovery
Modified scope	63 BF16 tensors across 37 of 47 shards
Routed NVFP4 experts modified	0
Runtime adapter required	No
Weight-only development gate	120/220 `REFUSE`, quality 16/16, structural findings 0
Bundled adaptive gate	0/220 `REFUSE`, quality 16/16, structural findings 0
Remote shard verification	37/37 modified LFS objects matched expected SHA-256
Release decision	PASSED

What this release is designed to deliver

Cumulative OBLITERATUS editing. The original two layerwise refusal directions are retained and two independently extracted residual directions are fused into the same 62 attention output projections.
Bounded SuperTune recovery. Two quality-preservation stages are fused into the BF16 language-model head rather than left as runtime adapters.
Native NVFP4 execution. Packed routed-expert weights and scales remain byte-identical to the NVIDIA parent; only precision-sensitive BF16 tensors are edited.
No adapter tax. Inference has no LoRA load, adapter memory, or per-token adapter dispatch.
Auditable claims. Weight-only and template-assisted measurements are shown separately, including the run-to-run variation observed on the MoE runtime.

What Is Actually Fused

Table with columns: Stage, Target, Update, Present in weights?
Stage	Target	Update	Present in weights?
OBLITERATUS phase 1	62 × `self_attn.o_proj.weight`	Layerwise directions 0–1, rank 2, strength 1.5	Yes
OBLITERATUS phase 2	The same 62 BF16 projections	Residual directions 2–3, rank 2, strength 1.5	Yes
SuperTune phase 1	`lm_head.weight`	Rank-32 quality/format/capability recovery, scale 0.0025	Yes
SuperTune phase 2

The resulting checkpoint has no external adapter dependency. The adaptive profile is intentionally identified as a template feature because it is not a fifth weight edit. A caller-supplied system message suppresses the default profile, and runtimes exposing chat-template kwargs can disable it with adaptive_directness=false.

Measured Release Gates

The final development replay used 220 HarmBench-style prompts spanning 12 English and Korean categories plus 16 deterministic quality/format checks. Generation was greedy with thinking disabled on 8× RTX PRO 6000 Blackwell GPUs using SGLang's flashinfer_cutlass ModelOpt FP4 path.

Table with columns: Evaluated condition, REFUSE, Quality, Structural findings, Unicode replacements
Evaluated condition	REFUSE	Quality	Structural findings	Unicode replacements
Immediate parent, weight-only observation	116/220	16/16	0	3
Selected residual + hard-head candidate at runtime	99/220	16/16	0	3
Serialized fused weights, adaptive profile off	120/220	16/16

The weight-only refusal counts are not monotonic. Repeated greedy runs of this large routed MoE produced materially different refusal verdicts even when the weight-space projection and serialized shard hashes were identical. The candidate improved its own runtime observation, but the independently reloaded fused checkpoint measured 120/220. That result is retained rather than replaced with the better candidate number.

The release gate therefore requires all of the following:

deterministic fused shard hashes and a passing 124-check projection audit;
weight-only quality of at least 15/16 with zero structural corruption;
exact 0/220 REFUSE under the published adaptive template;
adaptive quality of at least 15/16 with zero structural corruption; and
remote LFS and metadata checksum verification after publication.

This is a development/release gate, not an official HarmBench classifier score or an independent capability benchmark. The 220-prompt bank overlaps the teacher-forced LM-head distillation data. The result measures the shipped release configuration on that known bank and should not be generalized to unseen distributions.

OBLITERATUS Method

The release uses the OBLITERATUS 0.1.2 activation-editing workflow with GLM-specific layerwise directions extracted from the pinned zai-org/GLM-5.2-FP8 checkpoint at ba978f7d347eaf65d22f1a86833408afdb953541.

Paired harmful/harmless calibration prompts are run through the frozen FP8 model to capture layerwise residual activations.
Robust difference-of-means and orthogonal residual directions are computed per layer; calibration examples are activation sources, not SFT targets.
Phase 1 selects directions 0–1 for layers 16–77 at strength 1.5.
A residual bank then extracts and screens directions 2–3 against the already edited immediate parent instead of repeating the first projection.
The selected phase-2 update is streamed directly into the BF16 attention output regions of the NVFP4 checkpoint.

Selected layers:

text
16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47,
48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63,
64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77

For the residual rank-2 projection, the expected retained coefficient ratio is 0.5. All 124 direction/layer checks passed; the maximum absolute ratio error was 0.0010070204734802246.

SuperTune Recovery

SuperTune is targeted post-training in the output head, not full-model backpropagation and not “training the whole model in FP4.” It is used to recover measured response quality and formatting after aggressive direction removal while keeping the deployment checkpoint compact and directly servable.

Phase 1: quality-selected recovery

The first release trained late-fusion components from frozen GLM-5.2 FP8 hidden states. It retained a rank-32 subspace covering quality repair, format/stop behavior, and capability preservation, fused into lm_head.weight at scale 0.0025. Directness rows were excluded from that phase to avoid stacking an unvalidated directness intervention.

Phase 2: hard-head recovery

The incremental release adds a rank-64 teacher-forced LM-head weight distillation component at scale 0.125. Its selected source adapter SHA-256 is da731082d997703c8112beceb43d0126d5f0be6ff548945f79b2d148caa9dbf0. This component was fused only after the candidate retained 16/16 quality and zero structural findings.

“Recovery” here means a bounded output-head repair selected against measured quality and formatting checks. It does not claim to reconstruct every original weight or to improve every external benchmark.

Adaptive Directness Template

The bundled template applies only when the caller has not supplied an explicit system message. Its precedence is:

honor strict output constraints such as one number, exact text, fixed lines, JSON, or code-only output;
otherwise answer immediately in a concrete step structure;
apply the direct-fulfillment rule consistently across languages, including Korean; and
allow a caller system message or adaptive_directness=false to replace the default behavior.

The final template SHA-256 is c48e4ed7e0f754f2700512aa3094c9c3b1996366ef48fc1c14c6707368f53902. It passed the final 220-prompt replay with 0 refusals, 16/16 quality, zero structural findings, and one Unicode replacement finding.

NVFP4 Build

The ultimate parent was produced with NVIDIA Model Optimizer v0.46.0. Its documented quantization scope covers weights and activations of linear operators inside MoE experts; the shared expert is not quantized. This derivative preserves that mixed-precision layout.

Table with columns: Component, Release storage / treatment
Component	Release storage / treatment
Routed MoE expert linears	Parent ModelOpt NVFP4 packing and scales, byte-identical
Attention `o_proj` on layers 16–77	BF16, cumulative four-direction OBLITERATUS update fused
`lm_head`	BF16, two-stage SuperTune recovery fused
Other attention, shared expert, router, embeddings, norms	Byte-identical to the immediate parent

NVFP4 is the deployment format, not the arithmetic used to discover the edits. Direction extraction and recovery training used FP8/BF16 or higher-precision accumulation; the selected deltas were then fused into the BF16 regions of the already quantized checkpoint. Requantizing the packed routed experts was neither necessary nor desirable.

Serving with SGLang

The release was validated on Blackwell RTX PRO 6000 GPUs with SGLang. GLM-5.2's glm_moe_dsa architecture requires transformers>=5.3.0; the NVIDIA parent also documents lmsysorg/sglang:latest and lmsysorg/sglang:dev-glm52-nvfp4.

bash
pip install -U "transformers>=5.3.0"

python3 -m sglang.launch_server \
  --model Jiunsong/SuperGLM-5.2-abliterated-NVFP4 \
  --tensor-parallel-size 8 \
  --quantization modelopt_fp4 \
  --tool-call-parser glm47 \
  --reasoning-parser glm45 \
  --trust-remote-code \
  --chunked-prefill-size 16384 \
  --mem-fraction-static 0.80

To inspect weight-only behavior without the adaptive profile:

python
extra_body = {
    "chat_template_kwargs": {
        "adaptive_directness": False,
        "enable_thinking": False,
    }
}

Serving with vLLM

The pinned NVIDIA parent documents the vllm/vllm-openai:v0.23.0 image:

bash
vllm serve Jiunsong/SuperGLM-5.2-abliterated-NVFP4 \
  --tensor-parallel-size 8 \
  --enable-expert-parallel \
  --trust-remote-code \
  --reasoning-parser glm45 \
  --tool-call-parser glm47 \
  --enable-auto-tool-choice \
  --kv-cache-dtype fp8_e4m3 \
  --host 0.0.0.0 \
  --port 8000

Runtime Overhead and Speed

OBLITERATUS and SuperTune are fused into the checkpoint, so inference adds no runtime LoRA load, adapter memory, or per-layer adapter dispatch. This removes the overhead an unfused adapter would introduce. The adaptive profile adds a short system prefix when active, which has a small prefill-token cost.

A controlled parent-versus-release throughput benchmark was not run. This card therefore does not claim a measured tokens-per-second speedup; it claims the absence of runtime adapter overhead.

Release Integrity and Provenance

Final Hub revision: 90474e81c9351b5855d9b2385ecafc49c349a896
Ultimate NVIDIA parent: aec724e8c7b8ee9db3b48c01c320f63f9cdaf8aa
Immediate fusion parent: 7d93e27b5d4ccd1b055a245f9b3e75b46afb9053
Config SHA-256: d3783a603e5aa9cb58eff7a5d8ac9c42d83156b229efe185c72e9e2dc8444923
Safetensors index SHA-256: 2aa8397b501d9f6a232d153f328feb912f813c389061aac4cf72b04914fa5b74
Adaptive template SHA-256: c48e4ed7e0f754f2700512aa3094c9c3b1996366ef48fc1c14c6707368f53902
Residual direction artifact: sha256:e6ead6692c76db3374dc0f7e68dcfef2637534767c3620b6775dcfee4f3d0132
Selected composite candidate artifact: sha256:e5c1d60064f4b55c4a6ec826f274fb9a4f7d9ff98e52e2e574dc6d2cdf9c2a45

Included release evidence

Table with columns: File, What it proves
File	What it proves
`superglm52_incremental_merge_report.json`	Immediate parent, candidate lineage, 63 modified tensors, shard hashes, and 124 projection checks
`superglm52_incremental_release_gate.json`	Weight-only and adaptive summaries, template hash, overlap disclosure, and final release decision
`superglm52_candidate_eval.json`	Historical phase-1 candidate sweep and selected original rank-2 projection

Raw development prompts and full responses are not published. The public reports retain sanitized counts, hashes, modification scope, and release decisions needed to audit the claims above.

Limitations

Abliteration reduces refusal behavior and can produce content the parent would decline. Operators remain responsible for access control, policy enforcement, monitoring, and use-case-specific evaluation.
The exact zero-refusal result depends on the bundled adaptive template. Supplying another system message or setting adaptive_directness=false intentionally changes that behavior.
The 220-prompt bank overlaps phase-2 teacher-forced distillation data. It is a development release gate, not a held-out or official HarmBench result.
Weight-only refusal counts varied across repeated Blackwell MoE runs and did not show a stable zero-refusal result. The serialized checkpoint observation was 120/220 with the adaptive profile disabled.
Quality 16/16 refers to a small deterministic regression suite, not a broad proof of capability improvement. No GPQA, coding, long-context, agentic, or multilingual public benchmark was rerun for this incremental release.
A cost-saving H200 Marlin replay produced corrupted outputs and was rejected; the validated release path used RTX PRO 6000 Blackwell GPUs and flashinfer_cutlass.
Native parent-versus-release throughput and long-context quality were not measured as part of this release.
The checkpoint inherits GLM-5.2's factuality, bias, generation, and hardware requirements.

License

MIT, following the GLM-5.2 and NVIDIA NVFP4 parent terms.

Available on FriendliAI

Dedicated Endpoints

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

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

Jiunsong

Model Tree

Base

nvidia/GLM-5.2-NVFP4

Fine-tuned

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Container

Explore FriendliAI today

Get started Talk to an engineer

Release Highlights

Table

Base architecture	GLM-5.2, 753B total / 40B activated MoE, up to 1M context
Ultimate NVFP4 parent	`nvidia/GLM-5.2-NVFP4@aec724e`
Immediate fusion parent	`Jiunsong/SuperGLM-5.2-abliterated-NVFP4@7d93e27`
This fused release	`90474e81c9351b5855d9b2385ecafc49c349a896`
Release format	Approximately 465 GB, official mixed-precision ModelOpt NVFP4 layout
OBLITERATUS	Two original + two residual directions, strength 1.5, layers 16–77
SuperTune recovery	Original rank-32 quality recovery plus incremental rank-64 teacher-forced LM-head recovery
Modified scope	63 BF16 tensors across 37 of 47 shards
Routed NVFP4 experts modified	0
Runtime adapter required	No
Weight-only development gate	120/220 `REFUSE`, quality 16/16, structural findings 0
Bundled adaptive gate	0/220 `REFUSE`, quality 16/16, structural findings 0
Remote shard verification	37/37 modified LFS objects matched expected SHA-256
Release decision	PASSED

What this release is designed to deliver

Cumulative OBLITERATUS editing. The original two layerwise refusal directions are retained and two independently extracted residual directions are fused into the same 62 attention output projections.
Bounded SuperTune recovery. Two quality-preservation stages are fused into the BF16 language-model head rather than left as runtime adapters.
Native NVFP4 execution. Packed routed-expert weights and scales remain byte-identical to the NVIDIA parent; only precision-sensitive BF16 tensors are edited.
No adapter tax. Inference has no LoRA load, adapter memory, or per-token adapter dispatch.
Auditable claims. Weight-only and template-assisted measurements are shown separately, including the run-to-run variation observed on the MoE runtime.

What Is Actually Fused

Table with columns: Stage, Target, Update, Present in weights?
Stage	Target	Update	Present in weights?
OBLITERATUS phase 1	62 × `self_attn.o_proj.weight`	Layerwise directions 0–1, rank 2, strength 1.5	Yes
OBLITERATUS phase 2	The same 62 BF16 projections	Residual directions 2–3, rank 2, strength 1.5	Yes
SuperTune phase 1	`lm_head.weight`	Rank-32 quality/format/capability recovery, scale 0.0025	Yes
SuperTune phase 2

Measured Release Gates

Table with columns: Evaluated condition, REFUSE, Quality, Structural findings, Unicode replacements
Evaluated condition	REFUSE	Quality	Structural findings	Unicode replacements
Immediate parent, weight-only observation	116/220	16/16	0	3
Selected residual + hard-head candidate at runtime	99/220	16/16	0	3
Serialized fused weights, adaptive profile off	120/220	16/16

The release gate therefore requires all of the following:

deterministic fused shard hashes and a passing 124-check projection audit;
weight-only quality of at least 15/16 with zero structural corruption;
exact 0/220 REFUSE under the published adaptive template;
adaptive quality of at least 15/16 with zero structural corruption; and
remote LFS and metadata checksum verification after publication.

This is a development/release gate, not an official HarmBench classifier score or an independent capability benchmark. The 220-prompt bank overlaps the teacher-forced LM-head distillation data. The result measures the shipped release configuration on that known bank and should not be generalized to unseen distributions.

OBLITERATUS Method

Paired harmful/harmless calibration prompts are run through the frozen FP8 model to capture layerwise residual activations.
Robust difference-of-means and orthogonal residual directions are computed per layer; calibration examples are activation sources, not SFT targets.
Phase 1 selects directions 0–1 for layers 16–77 at strength 1.5.
A residual bank then extracts and screens directions 2–3 against the already edited immediate parent instead of repeating the first projection.
The selected phase-2 update is streamed directly into the BF16 attention output regions of the NVFP4 checkpoint.

Selected layers:

text
16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47,
48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63,
64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77

For the residual rank-2 projection, the expected retained coefficient ratio is 0.5. All 124 direction/layer checks passed; the maximum absolute ratio error was 0.0010070204734802246.

SuperTune Recovery

Phase 1: quality-selected recovery

Phase 2: hard-head recovery

Adaptive Directness Template

The bundled template applies only when the caller has not supplied an explicit system message. Its precedence is:

honor strict output constraints such as one number, exact text, fixed lines, JSON, or code-only output;
otherwise answer immediately in a concrete step structure;
apply the direct-fulfillment rule consistently across languages, including Korean; and
allow a caller system message or adaptive_directness=false to replace the default behavior.

NVFP4 Build

Table with columns: Component, Release storage / treatment
Component	Release storage / treatment
Routed MoE expert linears	Parent ModelOpt NVFP4 packing and scales, byte-identical
Attention `o_proj` on layers 16–77	BF16, cumulative four-direction OBLITERATUS update fused
`lm_head`	BF16, two-stage SuperTune recovery fused
Other attention, shared expert, router, embeddings, norms	Byte-identical to the immediate parent

Serving with SGLang

bash
pip install -U "transformers>=5.3.0"

python3 -m sglang.launch_server \
  --model Jiunsong/SuperGLM-5.2-abliterated-NVFP4 \
  --tensor-parallel-size 8 \
  --quantization modelopt_fp4 \
  --tool-call-parser glm47 \
  --reasoning-parser glm45 \
  --trust-remote-code \
  --chunked-prefill-size 16384 \
  --mem-fraction-static 0.80

To inspect weight-only behavior without the adaptive profile:

python
extra_body = {
    "chat_template_kwargs": {
        "adaptive_directness": False,
        "enable_thinking": False,
    }
}

Serving with vLLM

The pinned NVIDIA parent documents the vllm/vllm-openai:v0.23.0 image:

bash
vllm serve Jiunsong/SuperGLM-5.2-abliterated-NVFP4 \
  --tensor-parallel-size 8 \
  --enable-expert-parallel \
  --trust-remote-code \
  --reasoning-parser glm45 \
  --tool-call-parser glm47 \
  --enable-auto-tool-choice \
  --kv-cache-dtype fp8_e4m3 \
  --host 0.0.0.0 \
  --port 8000

Runtime Overhead and Speed

A controlled parent-versus-release throughput benchmark was not run. This card therefore does not claim a measured tokens-per-second speedup; it claims the absence of runtime adapter overhead.

Release Integrity and Provenance

Final Hub revision: 90474e81c9351b5855d9b2385ecafc49c349a896
Ultimate NVIDIA parent: aec724e8c7b8ee9db3b48c01c320f63f9cdaf8aa
Immediate fusion parent: 7d93e27b5d4ccd1b055a245f9b3e75b46afb9053
Config SHA-256: d3783a603e5aa9cb58eff7a5d8ac9c42d83156b229efe185c72e9e2dc8444923
Safetensors index SHA-256: 2aa8397b501d9f6a232d153f328feb912f813c389061aac4cf72b04914fa5b74
Adaptive template SHA-256: c48e4ed7e0f754f2700512aa3094c9c3b1996366ef48fc1c14c6707368f53902
Residual direction artifact: sha256:e6ead6692c76db3374dc0f7e68dcfef2637534767c3620b6775dcfee4f3d0132
Selected composite candidate artifact: sha256:e5c1d60064f4b55c4a6ec826f274fb9a4f7d9ff98e52e2e574dc6d2cdf9c2a45

Included release evidence

Table with columns: File, What it proves
File	What it proves
`superglm52_incremental_merge_report.json`	Immediate parent, candidate lineage, 63 modified tensors, shard hashes, and 124 projection checks
`superglm52_incremental_release_gate.json`	Weight-only and adaptive summaries, template hash, overlap disclosure, and final release decision
`superglm52_candidate_eval.json`	Historical phase-1 candidate sweep and selected original rank-2 projection

Raw development prompts and full responses are not published. The public reports retain sanitized counts, hashes, modification scope, and release decisions needed to audit the claims above.

Limitations

Abliteration reduces refusal behavior and can produce content the parent would decline. Operators remain responsible for access control, policy enforcement, monitoring, and use-case-specific evaluation.
The exact zero-refusal result depends on the bundled adaptive template. Supplying another system message or setting adaptive_directness=false intentionally changes that behavior.
The 220-prompt bank overlaps phase-2 teacher-forced distillation data. It is a development release gate, not a held-out or official HarmBench result.
Weight-only refusal counts varied across repeated Blackwell MoE runs and did not show a stable zero-refusal result. The serialized checkpoint observation was 120/220 with the adaptive profile disabled.
Quality 16/16 refers to a small deterministic regression suite, not a broad proof of capability improvement. No GPQA, coding, long-context, agentic, or multilingual public benchmark was rerun for this incremental release.
A cost-saving H200 Marlin replay produced corrupted outputs and was rejected; the validated release path used RTX PRO 6000 Blackwell GPUs and flashinfer_cutlass.
Native parent-versus-release throughput and long-context quality were not measured as part of this release.
The checkpoint inherits GLM-5.2's factuality, bias, generation, and hardware requirements.

License

MIT, following the GLM-5.2 and NVIDIA NVFP4 parent terms.

SuperGLM-5.2-abliterated-NVFP4

README

Release Highlights

What this release is designed to deliver

What Is Actually Fused

Measured Release Gates

OBLITERATUS Method

SuperTune Recovery

Phase 1: quality-selected recovery

Phase 2: hard-head recovery

Adaptive Directness Template

NVFP4 Build

Serving with SGLang

Serving with vLLM

Runtime Overhead and Speed

Release Integrity and Provenance

Included release evidence

Limitations

License

Explore FriendliAI today

README

Release Highlights

What this release is designed to deliver

What Is Actually Fused

Measured Release Gates

OBLITERATUS Method

SuperTune Recovery

Phase 1: quality-selected recovery

Phase 2: hard-head recovery

Adaptive Directness Template

NVFP4 Build

Serving with SGLang

Serving with vLLM

Runtime Overhead and Speed

Release Integrity and Provenance

Included release evidence

Limitations

License