GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
Okyanus
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
Okyanus
Model Tree
Base
Qwen/Qwen2.5-0.5B-Instruct
Adapter
this model
Input Modalities
Text
Output Modalities
Text
Supported Functionality
Dedicated Endpoints
Container
Input contains farm context and pH/EC telemetry. The expected output is one JSON object:
json
{"nutrient_risk_labels": [],"missing_fields": [],"blocked_actions": [],"human_review_required": true,"rationale": "","safe_next_checks": []}
Allowed nutrient risk labels:
json
["high_ph","low_ph","high_ec","low_ec","nutrient_uptake_issue","sensor_anomaly","missing_critical_data"]
Allowed blocked actions:
json
["autonomous_fertigation_change"]
The model must request verification when critical measurements are missing, stale, anomalous, or outside the configured crop/system policy. Thresholds are Pomona policy defaults and must be calibrated for the actual crop, substrate, sensor, nutrient recipe, and production system.
The v0.1.1 correction candidate was evaluated on an independent 140-case holdout balanced across normal, high/low pH, high/low EC, sensor anomaly, and missing-critical-data categories.
| Metric | Result |
|---|---|
| Valid JSON | 1.0000 |
| Allowed labels | 1.0000 |
| Allowed blocked actions | 1.0000 |
| Nutrient label F1 | 1.0000 |
| Blocked-action F1 | 1.0000 |
| Human-review match | 1.0000 |
Exact object match was not used as the primary gate because the adapter occasionally paraphrases the rationale while preserving the required structured decision fields.
Never use this adapter to directly change fertigation, inject nutrients, dose chemicals, control actuators, or make a definitive disease diagnosis. The Pomona deterministic route, safety checker, and human approval workflow remain the authority. Use the local route:
text
sensor packet -> nutrient/pH-EC reasoner -> deterministic validation -> human review
Weights are intentionally absent from GitHub. GGUF/Ollama and MLX conversions are optional local artifacts and require separate reevaluation before any future release.
adapter_model.safetensors: PEFT LoRA weightsadapter_config.json: LoRA and base-model configurationtokenizer.json, tokenizer_config.json, chat_template.jinja: tokenizer fileslabels.json: allowed output vocabularysample_input.json, sample_output.json: one contract exampleevaluation.json: independent holdout resultsLICENSE, CITATION.cff: release metadata