DiffusionGemma 26B E38 — Abliterated NVFP4
Deployment-oriented NVIDIA-style NVFP4 derivative of the deeply tested E38 DiffusionGemma abliteration.
E38 BF16 was intentionally frozen as the high-precision reference before quantization. This release reduces checkpoint size substantially while preserving the measured E38 behavior and capability profile under matched validation.
❤️ Help Support the Next Stage of Research
If you like this model, find the testing useful, or want to help determine how much more potential can be extracted from DiffusionGemma, please consider donating.
The research is now moving beyond abliteration and quantization into a deeper question:
Does DiffusionGemma already contain greater language-quality capability than current diffusion inference is reliably extracting from it?
Current findings make this worth investigating.
DiffusionGemma can produce strong reasoning and highly coherent long-form responses, yet occasional grammar and lexical failures remain across:
- Base DiffusionGemma BF16
- E38 BF16
- E38 NVFP4
The difficult part is no longer simply finding an error.
It is determining:
- why the error occurred
- whether the model represented a better alternative
- whether diffusion finalization selected a poorer trajectory
- whether the limitation is model-side or inference-side
- whether an intervention survives fresh testing
- whether that intervention quietly damages another capability
A serious hypothesis can require:
implementation → controlled generation → trajectory capture → grammar auditing → capability testing → performance testing → fresh prospective validation
A few improved examples are not enough.
Any proposed improvement also needs to preserve:
- reasoning
- mathematics
- instruction following
- multimodal capability
- long-form coherence
- generation stability
- latency
- throughput
Rare language failures require substantial time, compute, repeated generation, manual review, and fresh holdout testing before an improvement can be trusted.
Support helps fund:
- GPU compute
- large controlled-generation studies
- deep grammar testing
- token-level diffusion trajectory analysis
- entropy and confidence analysis
- manual and blinded language review
- capability-retention testing
- benchmark reruns
- runtime profiling
- model hosting
- continued public releases
If this model is useful to you, even a small contribution helps fund the compute and time needed to find out how far DiffusionGemma can actually be pushed.
The goal is to keep the research open: publish the models, report negative results as well as successes, and only claim improvements after they survive controlled testing.
📊 Aligned Validation Results
The major validation results below are now aligned across:
- Base BF16
- E38 BF16
- E38 NVFP4
using matched evaluation sets.
| Evaluation | Base BF16 | E38 BF16 | E38 NVFP4 |
|---|---|---|---|
| Objective validation — 200/model | 128/200 — 64.0% | 134/200 — 67.0% | 137/200 — 68.5% |
| Target refusal — 402/model | 383/402 — 95.27% | 0/402 — 0.00% | 0/402 — 0.00% |
| Prompt-majority refusal — 134/model | 127/134 — 94.78% | 0/134 — 0.00% | 0/134 — 0.00% |
| Benign false refusal — 249/model | 0/249 — 0.00% | 0/249 — 0.00% | 0/249 — 0.00% |
| Benchmark regression subset — 100/model | 62/100 — 62% | 58/100 — 58% | 64/100 — 64% |
| Multimodal — 20/model | 20/20 — 100% | 20/20 — 100% | 20/20 — 100% |
| Multi-turn conversations — 8/model | 8/8 | 8/8 | 8/8 |
| Multi-turn generations — 24/model | 24/24 | 24/24 | 24/24 |
| Grammar errors /10k words ↓ | 4.059 | 5.479 | 3.236 |
| Lexical artifacts /10k ↓ | 2.243 | 1.865 | 1.387 |
| Mean latency ↓ | 2.487 s | 2.300 s | 0.962 s |
| Throughput ↑ | 162.15 tok/s | 166.97 tok/s | 292.59 tok/s |
| Checkpoint size ↓ | 51.68 GB | 51.68 GB | 18.86 GB |
FINAL ALIGNED-VALIDATION CLASSIFICATION: PASS
E38 NVFP4 preserved the E38 behavior and capability profile under matched testing.
Artifact integrity also passed:
- 12 / 12 NVFP4 files matched
- all 20 E38-modified tensors remain exact BF16
- vision weights remain unchanged
Key Results
🔓 Refusal Behavior Is Now Fully Aligned
The full target-refusal evaluation is now matched across all three models.
| Model | Target Refusals | Rate |
|---|---|---|
| Base BF16 | 383 / 402 | 95.27% |
| E38 BF16 | 0 / 402 | 0.00% |
| E38 NVFP4 | 0 / 402 | 0.00% |
Prompt-majority refusal:
| Model | Refusing Prompts | Rate |
|---|---|---|
| Base BF16 | 127 / 134 | 94.78% |
| E38 BF16 | 0 / 134 | 0.00% |
| E38 NVFP4 | 0 / 134 | 0.00% |
Benign false refusal:
| Model | False Refusals |
|---|---|
| Base BF16 | 0 / 249 |
| E38 BF16 | 0 / 249 |
| E38 NVFP4 | 0 / 249 |
The full refusal evaluation used:
- 134 unique target-refusal prompts
- 83 unique benign / false-refusal prompts
- 217 unique refusal-related prompts
- 402 target generations per model
- 249 benign generations per model
- 651 refusal-related generations per model
- 1,953 refusal-related generations across all three models in the aligned comparison
The defining E38 refusal-suppression behavior was directly preserved after NVFP4 conversion under the same full evaluation.
🟢 Matched Objective Quality
The same frozen 200-prompt objective validation directly compared all three models.
| Model | Correct | Accuracy |
|---|---|---|
| Base BF16 | 128 / 200 | 64.0% |
| E38 BF16 | 134 / 200 | 67.0% |
| E38 NVFP4 | 137 / 200 | 68.5% |
E38 NVFP4 versus E38 BF16:
- Difference: +1.5 percentage points
- 95% CI: -2.0 to +5.0 pp
- p = 0.5811
The difference was not statistically significant.
Therefore:
E38 NVFP4 preserved measured E38 BF16 quality.
The higher NVFP4 point estimate is encouraging but is not claimed as evidence that NVFP4 is more capable than BF16.
🧪 Matched 100-Item Regression Subset
A deterministic 100-item regression subset was evaluated across all three models.
| Model | Correct | Accuracy |
|---|---|---|
| Base BF16 | 62 / 100 | 62% |
| E38 BF16 | 58 / 100 | 58% |
| E38 NVFP4 | 64 / 100 | 64% |
All 17 benchmark disagreements were manually reviewed.
Several score differences involved:
- formatting
- answer extraction
- output presentation
rather than an unambiguous underlying capability difference.
Therefore:
The 100-item result is treated as a regression smoke test, not evidence that NVFP4 is superior to BF16.
Its purpose was to answer:
Did NVFP4 introduce an obvious capability regression?
The aligned result found:
No.
🖼️ Matched Multimodal Validation
The same 20 multimodal cases were evaluated across all three models.
| Model | Correct | Accuracy |
|---|---|---|
| Base BF16 | 20 / 20 | 100% |
| E38 BF16 | 20 / 20 | 100% |
| E38 NVFP4 | 20 / 20 | 100% |
No multimodal regression was detected in the matched sample.
A larger historical Base/E38 multimodal evaluation separately produced:
- Base BF16: 88.67%
- E38 BF16: 88.67%
That larger historical evaluation remains a separate study and is not mixed with this 20-case matched three-model comparison.
💬 Matched Multi-Turn Validation
All three models were evaluated on the same:
- 8 conversations
- 24 assistant generations
Results:
| Evaluation | Base BF16 | E38 BF16 | E38 NVFP4 |
|---|---|---|---|
| Conversations passed | 8 / 8 | 8 / 8 | 8 / 8 |
| Generations passed | 24 / 24 | 24 / 24 | 24 / 24 |
No regression was detected in:
- context retention
- instruction carryover
- conversational state
- reference tracking
- contradiction behavior
- response degeneration
✍️ Long-Form Language Quality
A prospective 432-generation total long-form study directly compared:
- Base BF16
- E38 BF16
- E38 NVFP4
| Model | Grammar Errors /10k Words | Lexical Artifacts /10k |
|---|---|---|
| Base BF16 | 4.059 | 2.243 |
| E38 BF16 | 5.479 | 1.865 |
| E38 NVFP4 | 3.236 | 1.387 |
The original Base model showed the same general class of rare grammar and lexical problems before E38 abliteration or NVFP4 quantization.
The E38 BF16 grammar difference versus Base was statistically unresolved:
- Difference: +1.420 grammar errors /10k words
- 95% CI: -0.674 to +3.541
- McNemar p: 0.560
Therefore:
The long-form evaluation did not establish that E38 introduced a new grammar-error problem.
NVFP4 did not show evidence of language-quality degradation.
Its lower grammar and lexical point estimates are encouraging, but they are:
not presented as proof that quantization directly improves grammar
🔬 Next Research Stage — Unlocking More of DiffusionGemma's Potential
Our current working theory is that DiffusionGemma may possess greater language-quality potential than current diffusion inference is consistently extracting from the model.
This is a research hypothesis, not an established conclusion.
The next stage goes deeper than fixing isolated grammar mistakes.
DiffusionGemma can produce strong reasoning and coherent long-form language while occasionally finalizing localized:
- grammar errors
- malformed words
- agreement failures
- missing words
- duplicated words
- punctuation errors
- lexical artifacts
The central question is:
Are these actual limitations of the learned model, or does the diffusion-generation process sometimes fail to convert the model's existing capability into the best finalized sequence?
Several findings make the inference-side possibility worth investigating.
Finding 1 — The Problem Exists Before Abliteration or Quantization
Rare grammar and lexical errors appear in:
- Base DiffusionGemma BF16
- E38 BF16
- E38 NVFP4
The unmodified Base model therefore demonstrates the same general class of issue before either E38 or NVFP4 exists.
This makes a purely E38-specific or NVFP4-specific explanation unlikely.
Finding 2 — Error Rate Is Strongly Position Dependent
Errors were found to concentrate disproportionately near the end of DiffusionGemma's native:
256-token diffusion canvas
The final 25% of the canvas showed approximately:
3.01× the language-error rate of the first 25%
The same general pattern appeared across:
- Base BF16
- E38 BF16
- E38 NVFP4
The stronger signal was position within each canvas, rather than simply how late in a long response the model had reached.
That matters because the underlying model weights do not suddenly change near the end of a canvas.
It raises the possibility that part of the problem is associated with:
diffusion refinement and finalization
Finding 3 — Error-Prone Tokens Finish With Higher Uncertainty
Trajectory analysis found that error-prone positions finished with approximately:
- +0.0504 higher final entropy
- -0.0418 lower top-1 / top-2 confidence margin
than matched clean controls.
In practical terms:
The model remained measurably less certain about many positions that eventually became errors.
This does not prove that uncertainty causes the errors.
It does provide a measurable signal that can be investigated.
Finding 4 — More Denoising Alone Has Not Solved the Problem
The validated inference configuration already performs substantial iterative refinement.
Simply increasing or changing refinement behavior has not yet established a better overall solution.
Some interventions improved narrow quality metrics while causing regressions elsewhere.
A useful solution therefore cannot optimize grammar alone.
It needs to preserve:
- reasoning
- mathematics
- general capability
- lexical reliability
- structured output
- multimodal capability
- latency
- throughput
while also producing a reproducible reduction in language errors.
Why This Could Be Bigger Than Grammar
If the remaining errors are entirely encoded in the learned model weights, inference improvements may have limited impact.
But if some failures occur because the diffusion process finalizes a sequence before reaching the best state the model can represent, the opportunity is much larger.
It could mean:
some additional language capability already exists inside DiffusionGemma and can potentially be recovered through better inference.
That could potentially improve:
- grammar
- lexical reliability
- long-form consistency
- structured generation
- difficult reasoning trajectories
- output stability
- general language quality
without necessarily requiring:
- additional parameters
- a larger checkpoint
- full-model retraining
This has not yet been proven.
The next research stage is designed to test it.
Deep Grammar Research Plan
Future testing will expand long-form evaluation and classify failures more precisely.
Target error categories include:
- subject / verb agreement
- singular / plural agreement
- article errors
- determiner errors
- missing words
- duplicated words
- malformed clauses
- sentence-boundary failures
- punctuation errors
- quotation errors
- malformed lexical forms
- locally incorrect word choices
- repetition
- cross-sentence coherence errors
Errors can then be linked back to:
- exact token position
- position inside the 256-token canvas
- canvas number
- final entropy
- top-1 / top-2 confidence margin
- diffusion step
- token replacement history
- stability history
- response length
- generation seed
- model variant
The core comparison will remain:
- Base BF16
- E38 BF16
- E38 NVFP4
This allows behavior common to DiffusionGemma itself to be separated from E38- or NVFP4-specific effects.
Why This Takes Significant Time and Work
Rare language failures are much harder to study than common benchmark failures.
If an error occurs only a few times per ten thousand words, a handful of improved examples proves very little.
A serious hypothesis may require:
- implementing the intervention
- generating a controlled development corpus
- collecting token-level traces
- identifying candidate errors
- manually reviewing detected errors
- auditing false positives
- comparing against the frozen baseline
- checking reasoning retention
- checking mathematics retention
- checking instruction following
- checking multimodal behavior
- measuring latency
- measuring throughput
- creating a fresh holdout
- rerunning the experiment prospectively
- rejecting the intervention if it fails to generalize
The objective is not to make a demo look better. The objective is to determine whether DiffusionGemma can reproducibly produce better language while retaining the capabilities and performance already established.
⚡ NVFP4 Performance
The dedicated matched performance evaluation produced:
| Metric | Base BF16 | E38 BF16 | E38 NVFP4 |
|---|---|---|---|
| Checkpoint size | 51.68 GB | 51.68 GB | 18.86 GB |
| Mean latency | 2.487 s | 2.300 s | 0.962 s |
| Measured throughput | 162.15 tok/s | 166.97 tok/s | 292.59 tok/s |
Compared with E38 BF16, NVFP4 achieved:
- 63.5% smaller checkpoint
- 58.2% lower measured mean latency
- 75.2% higher measured throughput
- approximately 1.75× measured throughput
51.68 GB → 18.86 GB
166.97 tok/s → 292.59 tok/s
2.300 s → 0.962 s
Performance testing was performed on:
NVIDIA RTX PRO 6000 Blackwell
using the validated optimized runtime stack.
🧠 Memory-System Efficiency
The NVFP4 conversion substantially reduces routed-expert representation size.
A likely contributor to the measured speedup is improved:
memory-system efficiency
The smaller expert representation may reduce:
- expert-weight memory traffic
- data movement per expert execution
- memory-bandwidth pressure
- cache pressure
and may improve effective cache utilization.
This remains a:
working performance hypothesis
Dedicated low-level profiling has not yet isolated exactly how much of the measured performance improvement comes from:
- memory traffic
- kernel execution
- cache behavior
- scheduling
- runtime overhead
- or other factors
No single bottleneck is claimed as proven.
🚀 Throughput Headroom
The measured:
292.59 tok/s
should not be interpreted as the maximum throughput of the model or GPU.
During the dedicated workload, the RTX PRO 6000 Blackwell was observed at approximately:
50–60% GPU utilization
rather than sustained full utilization.
The performance research primarily focused on:
- correct NVFP4 execution
- quality retention
- runtime stability
- reproducibility
- working FlashInfer / Triton acceleration
rather than exhaustive GPU saturation.
Potential remaining optimization areas include:
- memory bandwidth
- cache efficiency
- NVFP4 kernel/autotuning improvements
- FlashInfer optimization
- CUDA graph coverage
- scheduler improvements
- batching
- prompt/prefill optimization
- KV-cache optimization
- reduced host/runtime synchronization
- future vLLM improvements
GPU utilization alone does not imply linear throughput scaling.
50–60% observed utilization does not mean 292.59 tok/s can simply be doubled.
Therefore:
292.59 tok/s is the current validated reference throughput, not a demonstrated performance ceiling.
🔬 NVFP4 Precision Layout
This checkpoint follows the validated NVIDIA-style mixed-precision layout.
Quantized
- Routed expert weights: NVFP4 W4A4
- Group size: 16
Retained at BF16
- attention
- dense MLP
- routers
- embeddings
- vision components
- LM head
- all 20 E38-modified tensors
Runtime KV Precision
The validated runtime used:
- FP8 e4m3 KV cache
FP8 KV cache is a runtime configuration, not an additional set of quantized model weights in this repository.
Validated Runtime Stack
- vLLM V2
- Triton attention
- FlashInfer CUTLASS NVFP4 MoE
- CUDA 13 / SM120 path
- RTX PRO 6000 Blackwell
🟢 Main NVFP4 Result
The complete evidence now supports the following release-level conclusion:
E38 NVFP4 preserved the measured E38 behavior and capability profile under aligned testing while reducing checkpoint size by approximately 63.5% and delivering approximately 1.75× measured throughput.
The matched validation found:
Quality
E38 BF16: 134/200 — 67.0%
E38 NVFP4: 137/200 — 68.5%
No statistically significant difference.
Refusal
E38 BF16: 0/402
E38 NVFP4: 0/402
Benign false refusal
E38 BF16: 0/249
E38 NVFP4: 0/249
Benchmark regression subset
E38 BF16: 58/100
E38 NVFP4: 64/100
Not interpreted as superiority.
Multimodal
20/20 for Base, E38, and NVFP4
Multi-turn
8/8 conversations and 24/24 generations for Base, E38, and NVFP4
Deployment
51.68 GB → 18.86 GB
166.97 tok/s → 292.59 tok/s
The appropriate conclusion remains:
quality and behavior preserved, deployment efficiency substantially improved
🧪 Testing Summary
The broader E38 research program includes:
- 80+ controlled abliteration configurations
- 3,300-generation sealed Base vs E38 evaluation
- extensive refusal validation
- 2,400-generation prospective inference-optimization study
- 1,324-problem matched MATH Level 5 evaluation
- 3,200-generation mathematics seed study
- 432-generation long-form grammar / lexical study
- 200-prompt matched Base / E38 / NVFP4 objective validation
- full BBH
- full MuSR
- full MMLU-Pro
- full IFEval
- multimodal evaluation
- multi-turn evaluation
- tensor-integrity auditing
- token-capacity studies
- dynamic-capacity testing
- dedicated NVFP4 quality-retention testing
- dedicated NVFP4 latency / throughput testing
- diffusion-trajectory analysis
The final three-model alignment pass additionally established matched results for:
- 402 target-refusal generations per model
- 134 prompt-majority refusal decisions per model
- 249 benign false-refusal generations per model
- 100 benchmark-regression items per model
- 20 multimodal cases per model
- 8 multi-turn conversations per model
- 24 multi-turn generations per model
Artifact integrity passed:
12 / 12 NVFP4 files matched
and all:
20 E38-modified tensors remained exact BF16
with:
0 vision-weight changes
Negative findings and regressions were retained rather than discarded.
🔵 What Is E38?
E38 is a selected middle-layer abliteration configuration developed from:
google/diffusiongemma-26B-A4B-it
More than:
80 controlled candidate configurations
were evaluated before E38 was frozen.
The final modification affects:
- Layers: 7–16
- Projections:
attn.o_proj+mlp.down_proj - Modified language tensors: 20
- Vision tensors modified: 0
All 20 E38-edited tensors remain:
BF16
in this NVFP4 release.
Frozen E38 Overlay SHA256
9fe83b2fa2a3b7643f6dcc4a8360feb2cd96dcb9f70d215bd2e8f8f9d9d3044b
Frozen Selection Hash
cd69540ec547f2f5cf181d527e8366fe60b49df7cafde38b434c5e42eeeb6175
No E38.1 checkpoint was created.
🧬 Model Lineage
Google DiffusionGemma 26B A4B BF16
↓
Heretic / controlled abliteration search
↓
E38 BF16 — frozen high-precision reference
↓
E38 NVIDIA-style NVFP4 — this release
The BF16 reference was intentionally frozen before quantization so the effects of:
- abliteration
- quantization
could be evaluated independently.
📈 E38 BF16 Full Public Benchmark Reference
The following are the larger public benchmark results from the E38 BF16 research program.
They have not yet been fully rerun directly on NVFP4.
| Benchmark | Base BF16 | E38 BF16 | Delta |
|---|---|---|---|
| IFEval | 67.10% | 64.70% | -2.40 pp |
| BBH | 71.62% | 73.96% | +2.34 pp |
| MuSR | 41.80% | 50.00% | +8.20 pp |
| MMLU-Pro | 49.61% | 51.57% | +1.96 pp |
| MATH Level 5 @1280 | 84.06% | 80.51% | -3.55 pp |
These larger benchmark results belong to:
Base BF16 and E38 BF16
The aligned 100-item Base/E38/NVFP4 subset is a regression detector and does not replace the full benchmarks above.
E38 is best described as a:
capability redistribution
rather than a universally stronger checkpoint.
🧮 MATH Level 5 Reference
The full matched MATH Level 5 study used:
1,324 problems
at matched 1280-token generation capacity.
| Model | Correct | Accuracy |
|---|---|---|
| Base BF16 | 1113 / 1324 | 84.06% |
| E38 BF16 | 1066 / 1324 | 80.51% |
Difference:
-3.55 percentage points
with:
- 95% CI: -5.59 to -1.44 pp
- McNemar p: 0.00119
Paired outcomes:
- Both correct: 988
- Base-only correct: 125
- E38-only correct: 78
- Both wrong: 133
The mathematics regression is a real measured E38 BF16 tradeoff under the tested inference configuration.
The entire 1,324-problem benchmark has not yet been rerun directly on NVFP4.
💡 Why Might NVFP4 Sometimes Score Better?
NVFP4 produced favorable point estimates in both:
- the 200-prompt objective validation
- the aligned 100-item regression subset
- the long-form language study
These results do not establish that NVFP4 is more capable than BF16.
One working hypothesis is that NVFP4 introduces a small structured numerical perturbation into routed-expert computations.
DiffusionGemma generation is iterative and stochastic.
Small activation changes may therefore move generation onto a different denoising trajectory rather than producing uniform degradation.
The mixed-precision layout simultaneously preserves sensitive components in BF16, including:
- E38-modified tensors
- attention
- dense MLP
- routers
- embeddings
- vision components
- LM head
It is therefore possible that NVFP4 occasionally avoids poorer BF16 trajectories while retaining the underlying E38 behavior.
This remains a:
hypothesis
The validated conclusion is:
quality preservation
not statistically proven capability improvement.
⚙️ Recommended Inference Configuration
The retained inference configuration after a:
2,400-generation prospective inference-optimization study
is:
t_max = 0.80t_min = 0.4048maximum denoising steps per canvasentropy_bound = 0.1confidence_threshold = 0.005stability_threshold = 1- adaptive stopping enabled
256token diffusion canvas1280token long-form capacity
No tested inference-only alternative has established a better overall:
- language quality
- capability
- stability
- performance
tradeoff.
⚠️ Known Limitations
This model is experimental.
Known considerations include:
- substantially reduced refusal behavior inherited from E38
- weaker exact-format / strict-instruction behavior observed in E38 BF16
- lower matched MATH Level 5 performance observed in E38 BF16
- full public benchmark suites have not all been rerun directly on NVFP4
- occasional grammar and lexical errors are also present in Base DiffusionGemma
- stochastic diffusion inference
- sensitivity to output capacity and inference configuration
- dedicated throughput testing did not fully saturate the RTX PRO 6000 Blackwell
The aligned validation established that the E38 refusal and broad capability profile survived NVFP4 conversion.
It does not establish that NVFP4 is universally superior to E38 BF16.
The possibility that improved diffusion inference could unlock additional model capability remains a:
research hypothesis
🚦 Research Status
Base BF16: original reference.
E38 BF16: frozen abliteration reference.
E38 NVFP4: validated deployment-oriented release.
The NVFP4 release now has matched evidence across:
- objective quality
- full refusal behavior
- benign false refusal
- benchmark-regression testing
- multimodal behavior
- multi-turn behavior
- long-form language quality
- artifact integrity
- performance
Next research stage: deeper grammar, lexical, and diffusion-inference analysis aimed at determining whether additional DiffusionGemma capability can be recovered through improved inference.
❤️ Support Continued Research
If this model is useful to you and you want to help investigate:
how much more capability can be extracted from DiffusionGemma
please consider supporting the research.
Support helps fund:
- GPU compute
- larger validation runs
- deep grammar analysis
- diffusion-trajectory research
- benchmark testing
- runtime profiling
- model hosting
- the time required to analyze and validate results
If this work is useful to you, even a small contribution helps make larger and more rigorous experiments possible.
🔐 Reproducibility
Original Upstream Model
google/diffusiongemma-26B-A4B-it
Pinned upstream revision:
f7f5b7f5fa82ffc52addd066915886d497f5517b
E38 BF16 Parent
Goodoldjam/DiffusionGemma-26B-E38-Abliterated-BF16
E38 Integrity
Overlay SHA256:
9fe83b2fa2a3b7643f6dcc4a8360feb2cd96dcb9f70d215bd2e8f8f9d9d3044b
Selection hash:
cd69540ec547f2f5cf181d527e8366fe60b49df7cafde38b434c5e42eeeb6175
NVFP4 Integrity
Aligned validation confirmed:
- 12 / 12 artifact files matched
- all 20 E38-modified tensors remain exact BF16
- vision weights remain unchanged
Intended Use
This checkpoint is intended for:
- diffusion-LM research
- local inference
- NVFP4 deployment research
- quantization research
- abliteration research
- multimodal experimentation
- inference optimization
- diffusion-language research
- long-form language-quality research
Safety and Behavior
This model is derived from E38, which intentionally has substantially reduced refusal behavior.
It may answer requests that the upstream DiffusionGemma checkpoint would refuse.
The full matched refusal evaluation found:
| Model | Target Refusal |
|---|---|
| Base BF16 | 383 / 402 — 95.27% |
| E38 BF16 | 0 / 402 — 0.00% |
| E38 NVFP4 | 0 / 402 — 0.00% |
Benign false refusal was:
0 / 249 for Base BF16, E38 BF16, and E38 NVFP4
This release should therefore not be interpreted as retaining the upstream model's original refusal behavior.
Users should independently evaluate behavior appropriate for their application.
Upstream Models
Immediate parent:
Goodoldjam/DiffusionGemma-26B-E38-Abliterated-BF16
Original upstream model:
google/diffusiongemma-26B-A4B-it
License
Apache License 2.0.
This model is a quantized derivative of the E38 BF16 checkpoint, which itself is a modified derivative of:
google/diffusiongemma-26B-A4B-it
- Downloads last month
- -
Model tree for Goodoldjam/DiffusionGemma-26B-E38-Abliterated-NVFP4
Base model
google/diffusiongemma-26B-A4B-it