Ekbasis-Skopos-27B-FP8
The FP8 build of Ekbasis-Skopos-27B (30.4 GB). Same model, prompt, readout, API and client; read its card for what the model is, its results and its limits.
FP8_DYNAMIC (FP8 weights per channel, FP8 activations computed per token at run time), made with llm-compressor; no calibration data. The vision tower, embeddings and LM head stay in bf16.
config.json declares mtp_num_hidden_layers: 1, but no multi-token-prediction weights are shipped: vLLM serves the build as it is; a loader or a speculative decoder that expects those weights will not find them.
Read the main card's limits before relying on this build in unfamiliar domains. In world families it never saw, Skopos gives more confidently wrong answers than Ekbasis-27B: 4.49 against 3.75 wrong answers at confidence 0.9 or more per 100 (+0.75 [+0.28, +1.23]; measured in bf16, not on this build).
Pre-launch evaluation (bf16; the main card's section of that name): one release blocker, declared and released by the owner's decision: short checks, paraphrased wording (120 items), 113 against Ekbasis-27B's 117 right (−3.3 points [−6.7, −0.8]; exact McNemar p = 0.125).
Quality against bf16
The release protocol on these exact files, served by vLLM and the System One API, as for the bf16 model:
| FP8 | Skopos bf16 | Ekbasis-27B | |
|---|---|---|---|
| git v3, known and held-out types, all questions | 96.88% | 96.94% | 96.62% |
| Never-trained worlds, changed answers (one prompt / read once) | 81.69% / 78.50% | 81.51% / 78.68% | 81.16% / 78.03% |
ftest_family, changed answers (629) |
80.45% | 80.29% | 81.88% |
| ECE on git v3 | 0.023 | 0.023 | 0.024 |
| Guard, known types: flagged / false alarms (of 863) | 835 / 27 | 837 / 27 | 828 / 37 |
| Guard, held-out types: flagged / false alarms (of 799) | 774 / 26 | 776 / 26 | 767 / 49 |
On the 3,014 changed answers of the fresh never-trained-family confirmation (5,385 items), this build (served) scores 83.2% against 81.8% for Ekbasis-27B in bf16 by trainer readout (the confirmation's reference; +1.4 points, 95% CI [+0.6, +2.3]). Only the summary of this run was kept, not its per-item answers.
The quantization gate of the Ekbasis repository (PREREG_quantized.md: accuracy within 1 point of bf16 on git, never-trained worlds and ftest_family, ECE within 0.01, the guard within 1 point, the same answer as bf16 on at least 97% of questions and 99% of its confident ones; applied here with the guard read on the larger guard sets above instead of git v3, and without the real-repository scenarios, T3c, which were not run on this build) is passed: same answer as bf16 on 98.57% of the questions and 99.95% of the answers bf16 gives at 0.9 or more.
decision_config.json sets this build's calibrated_temperature, 1.64, fit on validation (fval_git3 + m3val) for this build: serve.py then adds calibrated to every answer, as the hosted API does (remove the key to turn it off).
Quick start
hf download caiovicentino1/Ekbasis-Skopos-27B-FP8 --local-dir Ekbasis-Skopos-27B-FP8
bash Ekbasis-Skopos-27B-FP8/serve_vllm.sh Ekbasis-Skopos-27B-FP8 8001
python Ekbasis-Skopos-27B-FP8/serve.py --model Ekbasis-Skopos-27B-FP8 --vllm-url http://127.0.0.1:8001 --port 8000
pip install ekbasis && EKBASIS_URL=http://127.0.0.1:8000 ekbasis git-check -- "git reset --hard"
License
Apache-2.0, as Ekbasis-Skopos-27B (LICENSE, LICENSE-Qwen, LICENSE-Eikos, NOTICE). Its training data contains data
derived from ShIO-bash-26.1 (Ragsdale and Boppana, 2026) and swe-rebench-v2-CodeWorldModeling (Marin Community, 2026),
both CC BY 4.0: NOTICE gives the attributions and the changes made.