Coralfil/Atlantis-Pyramid-14B-Glyph-AWQ

Atlantis-Pyramid 14B (Glyph) — AWQ

Production 4-bit AWQ quantization of [`Coralfil/Atlantis-Pyramid-14B-Glyph`](https://huggingface.co/Coralfil/Atlantis-Pyramid-14B-Glyph) — Coralfil's marine-science LLM, purpose-built for the ocean economy and serving as the intelligence layer behind the Atlantis assistant in Coralfil OS (os.coralfil.com).

Atlantis-Pyramid (Glyph) is a domain-specialized model for marine biologists, aquaculture operators, coral and mariculture restoration teams, compliance and testing laboratories, and research educators. It reasons over ocean-climate stressors and produces field-actionable, science-first guidance — not generic chatbot answers.

Domain Focus

• —Ocean restoration — coral reef and oyster reef recovery planning, larval settlement enhancement, substrate and broodstock protocols
• —Ocean climate — marine heatwave / SST anomaly response, hypoxia and deoxygenation pathways, blue-carbon context
• —Ocean acidification remedies — aragonite saturation state (Ω) management, benthic pH buffering with mineral matrices (e.g., CaCO₃, shell-hash), slow-release alkalinity interventions
• —Aquaculture & mariculture — shellfish and finfish health, Vibrio spp. suppression, prebiotic formulation (FOS/XOS, chitin/chitosan, alginate carriers), feed and probiotic-vs-prebiotic decision support
• —Water management — salinity/nutrient/ turbidity interpretation, carrying-capacity reasoning for lease sites, integrated multi-trophic aquaculture (IMTA) balancing
• —Compliance & research — regulatory Q&A for testing labs, study design assistance for students and professors, formulation audit trails

Every formulation recommendation follows Coralfil's prebiotic-first doctrine: prebiotic interventions over probiotics, natural mineral/compound sources from marine waste streams, and explicit attention to pH, temperature, and pathogen suppression.

Release Notes — v2.0 (2026-08-10 UTC)

• —Renamed for parameter honesty. Formerly published as Atlantis-Pyramid-32B-AWQ; measured weights are ~14.7B parameters (Qwen2.5-14B dimensions: 48 layers, hidden 5120). The "32B" label was a lineage artifact — corrected across all repos for the Coralfil OS 2.0 launch. Old URLs redirect here permanently.
• —(Glyph) designation — this is the production Glyph-line release: the compact, fast, cost-efficient Atlantis tier designed for on-prem / single-GPU deployment without usage-meter burn.
• —First production quantized release. Replaces the interim GPT-2 124M placeholder that was mistakenly serving production chat (discovered during the Coralfil full-stack audit, 2026-08-09).
• —Quantized W4A16 (AWQ, GEMM, group size 128, zero-point) with AutoAWQ, default calibration.
• —Verified live on Lambda Labs A10 (24 GB): vLLM 0.26, --max-model-len 8192, API-key enforced (401 unauthenticated / 200 authenticated), engine --restart always.
• —End-to-end production verification (2026-08-10): live chat path os.coralfil.com → Cloudflare tunnel → FastAPI os-api → vLLM → Atlantis answered oyster-farm Vibrio/pH-drift and kelp-heat-stress prompts with structured, domain-correct formulation protocols.

Specification

Serving (vLLM)

docker run -d --gpus all --restart always -p 8080:8000 \
  -v /models:/workspace/models \
  vllm/vllm-openai:latest \
  --model /workspace/models/Atlantis-Pyramid-14B-Glyph-AWQ \
  --served-model-name atlantis-pyramid-14b-glyph-awq \
  --max-model-len 8192 \
  --gpu-memory-utilization 0.92 \
  --api-key "$VLLM_API_KEY"

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8080/v1", api_key="<VLLM_API_KEY>")
client.chat.completions.create(
    model="atlantis-pyramid-14b-glyph-awq",
    messages=[{"role": "user", "content": "Design a broodstock conditioning protocol for Crassostrea gigas under warming SSTs."}],
)

Deployment note: the current Coralfil production instance still answers to the legacy served id atlantis-pyramid-32b-awq; the id alias migrates to atlantis-pyramid-14b-glyph-awq with the OS 2.0 backend cutover.

Intended Use & Limitations

• —Intended for professional decision-support in marine science and aquaculture; outputs should be validated by qualified personnel before field deployment, regulatory submission, or animal-welfare-critical actions.
• —Trained for technical accuracy within marine domains; general-knowledge breadth is narrower than frontier general-purpose models.
• —License: other — see Coralfil for commercial terms.

Audit Trail

Full-stack audit lineage (naming correction F-34, placeholder-model discovery F-27, quantization + hardened serving MS-12) is documented in the Coralfil OS audit ledger. Base model training: SFT + DPO completed 2026-07-14→16 (wandb-tracked).