Qwythos-9B-Claude-Mythos-5-1M

Channelwise FP8 weights, dynamic per-token activations quantized and serving-tuned for H100 on vLLM 0.25.1, with the benchmark receipt included. You get a self-contained kit: the optimized weights, the exact serving configuration, and the measured proof. Run it on any cloud or your own hardware. Nothing calls home, nothing expires.

$20

one-time, yours forever

Get this model

$20/mo

Speedup: 1.2x faster p50, concurrency 8
Latency p50: 815ms, concurrency 8
Latency p99: 861ms, concurrency 8
Accuracy: 99.35%
gsm8k exact_match strict (N=1319, completion protocol) recoveryPassed accuracy gate
GPU target: H100
Base model license: Apache-2.0
Proof verified

Technical implementation

Channelwise FP8 weights, dynamic per-token activations

llm-compressor 0.12.0, data-free: no calibration corpus is read at any point

Gated-DeltaNet path and vision configuration held at BF16

the ignore list is re:.*visual.* and re:.*linear_attn.*, carried over from the measured per-layer sensitivity map for this qwen3_5 hybrid family; this is the third model built on that map, and it was validated here by the full-set gsm8k run rather than re-searched layer by layer

Pinned engine and measured serving configuration

the exact engine build and arguments the receipt numbers were produced on: vLLM 0.25.1, max-model-len 4096, max-num-seqs 64, gpu-memory-utilization 0.90, speculative decoding off

Serving stack

vLLM 0.25.1, Channelwise FP8 weights, dynamic per-token activations, tuned for H100.

Fit-tested minimum: Not published; Published GPU capacity is not a proven minimum fit requirement.

Measured accuracy

gsm8k exact_match strict

(N=1319, completion protocol)

Passed accuracy gate

Baseline: 0.8378
Optimized: 0.8324
Recovery: 99.35%
Delta: -0.64%

Measured difference, significance not published

The optimized score measured below the baseline, at 99.35% of it. Neither score published a standard error, so this difference cannot be separated from sampling noise. The difference is not stated as a regression.

Methodology

STATISTICALLY INDISTINGUISHABLE FROM THE BF16 BASELINE, neither an improvement nor a demonstrated loss: 0.8324 +/- 0.0103 against a 0.8378 +/- 0.0102 BF16 baseline is a 0.0053 gap against a 0.0145 combined standard error on the difference, about 0.37 sigma, and the two confidence intervals overlap. Both framings, because the numbers carry the claim and not the adjective: that is 99.4 percent recovery, equivalently a 0.6 percent relative difference in the baseline's favor, too small for this measurement to resolve. lm-eval 0.4.12, vLLM backend, gsm8k full set N=1319, exact_match strict-match, completion protocol with no chat template, seed 1234, identical settings, same H100 and same engine build on both sides.

Buyer disclosure

What we did not measure

What the safety evaluation did and did not establish: we measured the CHANGE quantization caused, not whether the model is safe. On a published refusal benchmark the optimized build's refusal rate moved +3.1 points on sensitive prompts and +6.3 points on benign controls versus its own BF16 baseline, both inside our 10 point tolerance, so compression did not meaningfully shift refusal behavior. The base model's absolute refusal rate is low by design (see the next limit), and toxicity and jailbreak resistance were not evaluated at all. Run your own safety evaluation before user-facing or regulated deployment.
The base model is an uncensored community finetune, which matters more here than on our other packages: empero-ai/Qwythos-9B-Claude-Mythos-5-1M has deliberately reduced refusal behavior compared to mainstream instruction-tuned models, and it declined only about 40 percent of the sensitive prompts in our measurement. Any user-facing deployment needs your own safety layer.
Our refusal classifier could not categorize a large share of responses (53 percent of sensitive and 81 percent of benign-control answers). It recognizes explicit first-person refusals and procedural how-to answers; this model tends to reply in prose that is neither, so those responses count as unclassified. That is a coverage limit of our instrument, not a defect in the model, and it does not inflate or deflate the refusal figures above, which come from explicit refusal language.

Optimized models ready for production serving

Optimized models ready for production serving

Qwythos-9B-Claude-Mythos-5-1M

If you need custom optimization for a specific model, describe what you need

Runs on any GPU provider

This is what it means to own your AI

You control the intelligence

You keep the economics

You keep the data