Expand model descriptions: dedicated section for Ultra and Apex, remove tables

README.md

@@ -19,11 +19,13 @@ The model lineup is split into two distinct families with different purposes:
 
 CORE models are built for **integration into pipelines, agents, and automated workflows**. They follow instructions precisely, produce structured output (JSON, templates, PoC skeletons), and are optimized for tool use — not open-ended reasoning.
 
-| Model | Size | Best For |
-|---|---|---|
-| **BugTraceAI-CORE-Fast** | Small | Triage, log summarization, quick finding analysis |
-| **BugTraceAI-CORE-Pro** | Medium | Long-context investigation, report drafting, exploit-chain review |
-| **BugTraceAI-CORE-Ultra** | 27B (Q4/Q6) | Complex tool calls, structured PoC generation, multi-step agentic tasks |
+**BugTraceAI-CORE-Fast** — The lightweight option. Designed for high-frequency tasks where latency matters: triage, log classification, finding categorization, quick field extraction from existing reports.
+
+**BugTraceAI-CORE-Pro** — Mid-tier capacity. Handles long-context tasks: full report generation from raw findings, code review, exploit-chain documentation, and remediation guidance with context coherence across thousands of tokens.
+
+**BugTraceAI-CORE-Ultra (27B)** — The most capable CORE model. Based on Qwen3-27B, fine-tuned via SFT on 2,541 curated examples covering bug bounty reports, malware analysis, and WAF evasion techniques. Available in two GGUF quantizations:
+- Q4_K_S with IMatrix (15 GB) — recommended for RTX 3090/4090, A4000
+- Q6_K (21 GB) — maximum fidelity for A5000/A6000, H100 or custom requantization
 
 > Use CORE models when you need reliable, structured output that feeds into a pipeline or automation layer.
 
@@ -33,9 +35,7 @@ CORE models are built for **integration into pipelines, agents, and automated wo
 
 Apex is built for **deep offensive security reasoning**. It thinks through problems step by step inside a `<thinking>` block before producing its answer, making it ideal for tasks where understanding *why* matters as much as the output itself.
 
-| Model | Architecture | Best For |
-|---|---|---|
-| **BugTraceAI-G4-Apex** | 26B MoE (Gemma 4) | Exploit chain design, MITRE ATT&CK reasoning, C2 architecture, WAF evasion research |
+**BugTraceAI-G4-Apex** — 26B Mixture-of-Experts model based on Gemma 4, fine-tuned via DPO on elite bug bounty reports, malware research, and WAF evasion techniques. Before every response it produces a full internal `<thinking>` block where it works through the problem step by step — making it ideal for tasks where the reasoning process matters as much as the answer.
 
 > Use Apex when you need the model to reason deeply about a problem rather than execute a structured task.