GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
ccortezb
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
ccortezb
Model Tree
Qwen/Qwen2.5-14B-Instruct
Input Modalities
Output Modalities
Supported Functionality
v1 was trained for a stack that has since evolved. v2 re-aligns the model to the current ecosystem and adds three new domains:
| Capability | v1 | v2 |
|---|---|---|
| Terraform HCL audit (rightsizing, tagging, Graviton, gp2→gp3, lifecycle) | ✅ | ✅ updated |
| AWS Cost Explorer JSON + anomalies | ✅ | ✅ |
| Savings Plans / Reserved Instances coverage & utilization | — | ✅ new |
| BYaML v2 Architecture Graph reasoning (blast radius, cost chains, owner attribution) | YAML lint only | ✅ graph-first |
| Remo — incident triage + remediation-as-code | — | ✅ new |
| sofe-dbt — data-engineering FinOps (Athena bytes scanned, incremental vs full-refresh, partitioning) | — | ✅ new |
| FinOps Q&A | ✅ | ✅ |
Key shift: BYaML moved from "validate this YAML" to "reason over the derived Architecture Graph" — the model computes blast radius and cost propagation over nodes (aws.<resource>) and typed edges, matching the real byaml-spec v0.4 schema.
| Parameter | Value |
|---|---|
| Base model | Qwen/Qwen2.5-14B-Instruct (4-bit via Unsloth) |
| Method | QLoRA (r=16, α=32, dropout=0.05) |
| Framework | Unsloth + TRL (SFTTrainer) |
| Dataset | 966 examples (265 v1 + 701 new, deterministic generators) |
| Hardware | Modal A100-80GB |
| Duration | ~20 minutes, 3 epochs, 183 steps |
| Final train loss | 0.28 |
| Cost | ~$2 |
The dataset was built with deterministic per-domain generators faithful to the real policy definitions (sofe-dbt policies, SOFE savings collectors, the byaml-spec v0.4 schema) — every cost figure is derived from the input, never fabricated.
python
from transformers import AutoModelForCausalLM, AutoTokenizer model = AutoModelForCausalLM.from_pretrained("ccortezb/FinOptix-14B-v2", device_map="auto")tok = AutoTokenizer.from_pretrained("ccortezb/FinOptix-14B-v2") messages = [ {"role": "system", "content": "You are FinOptix-14B, a FinOps specialist. Costs are derived from real data, never invented."}, {"role": "user", "content": "Audit this dbt model: fct_events runs daily with --full-refresh, bytes_scanned_last_run=1.2TB, is_incremental=1."},]inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)out = model.generate(inputs, max_new_tokens=350, do_sample=False, repetition_penalty=1.1)print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))bash
ollama run hf.co/ccortezb/FinOptix-14B-v2-GGUF:Q4_K_M# or higher fidelity:ollama run hf.co/ccortezb/FinOptix-14B-v2-GGUF:Q6_KRequires ~10GB RAM/VRAM (Q4_K_M) or ~13GB (Q6_K).
dbt cost audit → detects --full-refresh on a 1.2TB incremental model, recommends partition_by + removing the daily full-refresh (~$180/mo Athena savings), classifies as Optimize.
Incident triage (Remo) → classifies severity, finds root cause, and emits remediation-as-code (e.g. Terraform aws_s3_bucket_public_access_block for a recurring public bucket).
Architecture Graph blast radius → given a BYaML v0.4 graph, computes which nodes break if a node fails and which cost chains drive spend.
Full 3-part build log: BLOG.md in this repo.
Apache-2.0. Built by Carlos Cortez — AWS Community Hero. LinkedIn · X · GitHub · AWS Heroes
from transformers import AutoModelForCausalLM, AutoTokenizer model = AutoModelForCausalLM.from_pretrained("ccortezb/FinOptix-14B-v2", device_map="auto")tok = AutoTokenizer.from_pretrained("ccortezb/FinOptix-14B-v2") messages = [ {"role": "system", "content": "You are FinOptix-14B, a FinOps specialist. Costs are derived from real data, never invented."}, {"role": "user", "content": "Audit this dbt model: fct_events runs daily with --full-refresh, bytes_scanned_last_run=1.2TB, is_incremental=1."},]inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)out = model.generate(inputs, max_new_tokens=350, do_sample=False, repetition_penalty=1.1)print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))ollama run hf.co/ccortezb/FinOptix-14B-v2-GGUF:Q4_K_M# or higher fidelity:ollama run hf.co/ccortezb/FinOptix-14B-v2-GGUF:Q6_K