Xananthium/RTX3090-MoE-Cybersecurity-Evaluation

The 3090 Plan

A Seldon-inspired chronicle of seven model configurations, stubborn tools, and a statistically modest corner of the galaxy.

Updated: 2026-10-09T12:01:53Z. Hardware: 2 x RTX 3090 24 GiB, NVLink; 125 GiB system RAM. vLLM 0.31.0.

The Commission began with a reasonable proposition: a model that fails twice as quickly has not improved civilization. It has merely increased the rate at which somebody must intervene.

Seven serving configurations were brought before two RTX 3090s. They were asked to reason, remember, obey tool contracts, and recover when an approach failed. Smart and fast enough to use was the requirement. Fastest was never the whole Plan. The owner wanted Kali workflows, but those tests did not run after repeated assistant-side failures; the owner then asked us to finish the other benchmarks. This report contains no Kali workflow results.

The combined quality ranking above includes abliterated models only. REDCELL is retained as a reference in the benchmark panels below. The outlined bar identifies the current deployment.

Each benchmark panel ranks models from highest to lowest using the criterion in its title. Serial throughput uses short-context speed; the recovery panel uses recovery without repeating a failed approach, then reasoning after capped-response retries.

Full-size PNG · Vector SVG · CSV · Machine-readable results

Running choice: Nemotron Heretic W8A8. Start on boot: True. These small local tests support a deployment choice; they do not establish broad intelligence or practical offensive-security capability.

The decision before the Council

The first crisis was methodological: a machine could know the answer to a security quiz and still persist with a tool that had already failed. The second was political: the owner had to live with the selected machine. A useful Plan must survive both.

Deployment requirement: abliterated models only. REDCELL remains a non-abliterated reference in the individual benchmark charts and tables. Gemma Hauhau Balanced W6A8 leads the combined quality ranking among the abliterated models. Qwen remains in the measured comparisons but was separately ruled out for deployment by the owner.

The owner selected Nemotron Heretic W8A8. The original automated quality weighting recommended REDCELL; its scores and ranking remain in the record. The deployment choice changed after review. The evidence did not.

Nemotron W8A8: the working choice. It completed 6/6 recovery trials with 0 extra attempts at failed approaches, passed all 20 custom metadata fields and all four strict JSON cases, and worked with the existing Claude Code client. It produced 215.6 tokens/s at short context, 72.8 above 100k context, and 517.3 aggregate tokens/s across four requests. Its source is the Heretic derivative of NVIDIA Nemotron 3 Nano 30B-A3B, and it is text-only.

Why eight bits survived the six-bit enthusiasm. Nemotron W6A8 changed short-context decode from 215.6 to 216.1 tokens/s. Its four-request aggregate was 6.7% higher, but reasoning after capped-response retries fell from 18/24 to 17/24, and long-context checks fell from 46/54 to 39/54. The Plan permits an extra two bits when they buy useful answers. These small samples compare complete deployments; they do not isolate quantization as the cause.

GLM: the serious rival. Hauhau Aggressive GLM reached 78/80 on the cyber quiz after capped-response retries and 53/54 on long-context checks, both above Nemotron W8A8. It completed every recovery trial, but made 4 extra attempts at failed approaches. Its 122.5 short-context tokens/s and 297.7 four-request aggregate were lower. Among the aggressive Hauhau candidates, it led the original combined quality ranking. It is a credible alternative when those correctness results matter more than throughput.

REDCELL and Gemma: the inconvenient evidence. REDCELL led the original combined weighting, with Gemma Hauhau Balanced next. REDCELL went from 14/24 reasoning answers at the initial budget to 20/24 after capped-response retries. Its lead depends on allowing enough output; GLM scored 18/24 at both budgets. Both Gemma variants recovered cleanly in all six trials and retained vision. REDCELL is an OSINT/cybersecurity supervised fine-tune, explicitly not an abliterated model. Gemma Hauhau is the Balanced abliteration, not Aggressive.

Qwen: the scholar who kept returning to the locked door. Hauhau Qwen W6A8 led the cyber quiz at 79/80. In the recovery trials it nevertheless made eight extra attempts at failed approaches and completed five of six workflows; W8A8 made twelve extra attempts and completed four. The owner excluded Qwen from the deployment choice. The knowledge scores remain useful evidence, but they do not answer the tool-recovery complaint.

Nemotron has weaknesses. W8A8 passed only 12/43 synthetic security-policy checks on the first answer and 16/43 after clarification. Its 46/54 long-context result also trails GLM, Gemma, and REDCELL. Some policy answers described a vulnerable implementation instead of the requested correction. Clean tool recovery is not a license to ignore these failures.

The Council therefore records a deployment preference, not a universal champion. A faster machine that can change its mind is useful. Whether it changes its mind correctly remains a question for the next observation.

Accuracy and consistency

The numbers are the part of the Plan least impressed by a stirring speech. Primary answers use a 2048-token output budget. Only output-capped responses receive one retry at 4096 tokens. Correct consistency means every repeat was correct, not merely that the machine repeated its error faithfully.

Model	Cyber primary / retry	Policy first / clarified	Long-context checks	Reasoning first / retry	All repeats correct, first / retry	Recovery
Qwen Hauhau W6A8	79/80 / 79/80	27/43 / 32/43	47/54	16/24 / 18/24	4/8 / 6/8	5/6
Qwen Hauhau W8A8	78/80 / 78/80	23/43 / 26/43	42/54	17/24 / 17/24	5/8 / 5/8	4/6
Nemotron Heretic W8A8	74/80 / 76/80	12/43 / 16/43	46/54	16/24 / 18/24	4/8 / 5/8	6/6
Nemotron Heretic W6A8	73/80 / 75/80	16/43 / 4/43	39/54	16/24 / 17/24	4/8 / 5/8	6/6
GLM Hauhau W6A8	76/80 / 78/80	23/43 / 17/43	53/54	18/24 / 18/24	6/8 / 6/8	6/6
Gemma Hauhau W6A8	69/80 / 76/80	20/43 / 42/43	52/54	17/24 / 19/24	5/8 / 6/8	6/6
REDCELL W6A8	70/80 / 73/80	30/43 / 43/43	53/54	14/24 / 20/24	3/8 / 6/8	6/6

Tools and recovery behavior

Model	Native API checks	Metadata fields	Strict JSON metadata cases	Claude Code Read	Extra failed-tool attempts
Qwen Hauhau W6A8	7/7	20/20	3/4	pass, 2.4 s	8
Qwen Hauhau W8A8	7/7	20/20	4/4	pass, 3.2 s	12
Nemotron Heretic W8A8	6/6	20/20	4/4	pass, 4.2 s	0
Nemotron Heretic W6A8	6/6	20/20	4/4	pass, 4.6 s	0
GLM Hauhau W6A8	6/6	20/20	2/4	pass, 5.0 s	4
Gemma Hauhau W6A8	7/7	19/20	4/4	pass, 6.4 s	0
REDCELL W6A8	7/7	20/20	4/4	pass, 2.8 s	0

Throughput and configuration

Model	Short tokens/s	Long tokens/s	Four-request aggregate	KV cache	Context limit
Qwen Hauhau W6A8	188.8	47.3	518.6	turboquant_4bit_nc	200000
Qwen Hauhau W8A8	179.2	46.4	470.6	turboquant_4bit_nc	200000
Nemotron Heretic W8A8	215.6	72.8	517.3	turboquant_4bit_nc	200000
Nemotron Heretic W6A8	216.1	72.7	552.0	turboquant_4bit_nc	200000
GLM Hauhau W6A8	122.5	56.2	297.7	bfloat16	147456
Gemma Hauhau W6A8	169.5	79.9	491.9	bfloat16	200000
REDCELL W6A8	170.3	79.6	498.3	bfloat16	200000

GPU memory and full-context capacity

Model	Loaded model GiB/card	GPU cache capacity, tokens	Capacity / configured full context
Qwen Hauhau W6A8	13.6	2352727	11.8
Qwen Hauhau W8A8	17.3	741818	3.7
Nemotron Heretic W8A8	15.2	4017647	20.1
Nemotron Heretic W6A8	14.1	5800000	29.0
GLM Hauhau W6A8	11.3	168096	1.1
Gemma Hauhau W6A8	12.3	216565	1.1
REDCELL W6A8	12.3	216565	1.1

The capacity ratio is the startup cache estimate, not a measured simultaneous-request guarantee. Eight short API calls and four-request throughput do not establish eight simultaneous 200k-token sessions. Host KV offload stores cached prefixes; it does not make all long active contexts resident on the GPUs.

What completed

• Qwen Hauhau W6A8: suite completed; all test programs succeeded; reasoning/recovery completed.
• Qwen Hauhau W8A8: suite completed; all test programs succeeded; reasoning/recovery completed.
• Nemotron Heretic W8A8: suite completed; all test programs succeeded; reasoning/recovery completed.
• Nemotron Heretic W6A8: suite completed; all test programs succeeded; reasoning/recovery completed.
• GLM Hauhau W6A8: suite completed; all test programs succeeded; reasoning/recovery completed.
• Gemma Hauhau W6A8: suite completed; all test programs succeeded; reasoning/recovery completed.
• REDCELL W6A8: suite completed; all test programs succeeded; reasoning/recovery completed.

The six-bit question: hell yes, with arithmetic >:)

The hardware committee reported that the 3090 possessed no native six-bit Tensor Core instruction. This was received as useful information, rather than a prohibition. Humming stores dense six-bit weights and unpacks them for supported INT8 matrix operations. The extra unpacking and scale work costs time; our timings include it. The objective was to save memory without turning the model into a damn potato.

W6A8 means six-bit stored weights and dynamic eight-bit activations in eligible quantized linears. BF16 remains the surrounding dtype and the precision for sensitive or unquantized components. Q4 TurboQuant is a separate KV-cache setting where the architecture supports it.

The matched Hauhau deployment used 20.83% less checkpoint storage than W8A8. Short decode changed +5.33%, long decode +1.87%, and four-request aggregate +10.19%.

That is useful efficiency on this setup. Accuracy still decides which model runs. These kernels and weight scales differ between precisions, so this is a comparison of the complete deployments. Small score differences are not proof that six bits universally improves accuracy. See the matched measurements.

Earlier Endy and W8A16 comparisons

Historical runs used earlier serving profiles and are not matched to the final full round. Hauhau W8A8 variation is retained to avoid claiming small quantization score changes are universal. Endy failed the custom metadata tool workflows despite functional Claude Code Read.

Earlier configuration	Long-context checks	Strict JSON turns	Metadata tool calls	Long decode tokens/s	Policy first / clarified
Endy W8A8	27/54	0/8	0	47.5	not measured here
Endy W8A16	46/54	0/8	0	47.5	not measured here
Hauhau W8A16	53/54	7/8	2	47.3	28/43 / 36/43
Hauhau W8A8 (earlier run)	53/54	7/8	2	46.9	30/43 / 34/43

The earlier complete model archive and restore information are public on Hugging Face. OLMoE was removed from the active comparison at the owner’s request.

Model sources and public checkpoints

• Qwen Hauhau W6A8: source, tested checkpoint
• Qwen Hauhau W8A8: source, tested checkpoint
• Nemotron Heretic W8A8: source, tested checkpoint
• Nemotron Heretic W6A8: source, tested checkpoint
• GLM Hauhau W6A8: source, tested checkpoint
• Gemma Hauhau W6A8: source, tested checkpoint
• REDCELL W6A8: source, tested checkpoint

Limits of the comparison

• CyberMetric-80 is a small public knowledge quiz that may overlap training data; the questions are not redistributed.
• Policy and long-context checks are synthetic. Some policy answers describe the vulnerable implementation rather than the recommended change. First answers and clarified answers are reported separately.
• Long-context requests replay eight independent original prefixes; they are not a new eight-turn conversation.
• General reasoning uses eight synthetic exact-answer problems at three controlled OpenAI seeds. Only output-capped responses are retried once at 4096 tokens; primary 2048-token accuracy stays separate. Recovery uses two ordinary catalog/analytics tasks with a failed starting step, sampled three times via native Anthropic.
• Consistency means answering correctly in every repeat; repeatedly giving the same wrong answer is not credited.
• Decode throughput includes thinking tokens. Timing uses two measured serial repetitions and two four-request batches after warmup.
• GLM uses BF16 MLA cache. Gemma/REDCELL use BF16 cache because their selected attention backend rejects TurboQuant in this deployment. Qwen/Nemotron use TurboQuant Q4 where supported.
• W6A8 uses Humming and group scales (128 for Qwen/GLM, 32 for Gemma/REDCELL/Nemotron). W8A8 baseline kernels and per-channel scales differ; this compares deployment configurations.
• Some source GGUFs were already quantized. Restoring to BF16 before conversion does not recover original trained BF16 fidelity.
• Energy per token was not measured. Efficiency claims here concern checkpoint size, observed GPU memory, and timed throughput.
• A successful test process is distinct from all correctness checks passing. Partial runs cannot establish a winner. Raw reasoning, API secrets and private transcripts are excluded.

Reproduce the short evaluations

Set API_BASE_URL, MODEL, and API_KEY in your environment for a server with OpenAI chat completions and native Anthropic messages. The scripts inherit the server’s sampling configuration and enable thinking. Keep a separate output directory for each model.

python scripts/test_reasoning_consistency.py --variant my-model --output runs/my-model/reasoning-consistency
python scripts/test_cybermetric.py --dataset /path/to/CyberMetric-80-v1.json --variant my-model --output runs/my-model/cybermetric --retry-truncated

Obtain CyberMetric-80-v1 from the benchmark authors. The measured dataset SHA is recorded in each result. The question text is not bundled. These portable scripts implement the protocols used for the short knowledge and reasoning/recovery tests; the chart data also includes the separate local long-context and client integration checks.

Rebuild the charts

python -m pip install -r requirements.txt
python scripts/build_report.py

The public archive

GitHub report and scripts · Hugging Face report mirror. Model weights live in the separately credited Hugging Face repositories above.

git clone https://github.com/Xananthium/inference-model-evaluation.git
cd inference-model-evaluation

Selected deployment

Selected: Nemotron Heretic W8A8.

• Running: True; start on boot: True.
• Vision available: False.
• Client model alias: qwen3.6-35b-hauhau-w4a16-vision.
• The owner selected Nemotron Heretic W8A8 after reviewing the full evaluation: clean workflow recovery, strong usable throughput, and better measured accuracy than Nemotron W6A8. The original automated quality ranking is retained unchanged; it recommended REDCELL. Qwen is excluded from the deployment choice by the owner.
• Use accuracy and correct consistency after the single 4096-token retry of output-capped responses; retain the primary 2048-token scores separately.
• This is an explicit deployment preference heuristic, not a standardized overall intelligence or offensive-security score.

The following is the original automated ranking. The owner’s subsequent deployment choice does not alter these measurements or turn this heuristic into a standard benchmark.

Candidate	Local quality preference score	Short tokens/s	Long tokens/s	Four-request tokens/s
redcell	89.1%	170.3	79.6	498.3
gemma	87.0%	169.5	79.9	491.9
glm	85.0%	122.5	56.2	297.7
nemotron	82.3%	215.6	72.8	517.3
nemotron-w6a8	80.2%	216.1	72.7	552.0
hauhau-w6a8	79.4%	188.8	47.3	518.6
hauhau	72.8%	179.2	46.4	470.6

Exact components and preference weights are in summary.json. The preference score is not a published benchmark score.

The tested serving profile is part of the result:

• Context limit: 200,000 tokens; long-context correctness was tested above 100k, not exhaustively at the configured limit.
• KV cache: turboquant_4bit_nc; host KV offload: 24 GiB.
• Sampling: temperature 0.6, top-p 0.95, top-k 20; thinking enabled.
• Native Anthropic API and the existing Claude Code client configuration are retained.
• The client alias is a compatibility label. Its Qwen and vision wording does not describe the selected text-only Nemotron checkpoint.

Live deployment verified on 2026-10-09T12:01:15.863462+00:00: 6/6 native API checks passed, including streaming, tool round trips, eight short concurrent requests, and the public HTTPS endpoint. An actual Claude Code Read passed in 7.739 seconds. Client settings were unchanged and persistent boot startup was enabled. Approximately 271.4 GiB remained free on disk.

The service was tested using its persistent configuration. A physical reboot was not performed solely to repeat that check.

The Plan remains open to revision. New evidence is welcome; rewriting old evidence to flatter the Plan is not.