README: add SGLang config + real performance numbers

README.md

@@ -37,20 +37,51 @@ security workflows and autonomous security agents.
 | Tool-calling | OpenAI-compatible (`qwen3_coder` parser) |
 | Serving | SGLang / vLLM compatible |
 
-## Usage
-Nesi1 runs on any OpenAI-compatible inference server (SGLang, vLLM). Example
-with SGLang:
+## Deployment
+Nesi1 runs on any OpenAI-compatible inference server (SGLang, vLLM). This is
+the exact SGLang configuration we run in production:
 
 ```bash
 python3 -m sglang.launch_server \
   --model-path nesilabs/Nesi1 \
   --served-model-name Nesi1 \
   --context-length 262144 \
+  --mem-fraction-static 0.92 \
+  --trust-remote-code \
+  --kv-cache-dtype auto \
   --tool-call-parser qwen3_coder \
-  --trust-remote-code
+  --default-chat-template-kwargs '{"enable_thinking":false}' \
+  --host 0.0.0.0 --port 30000
 ```
 
-Then query it like any OpenAI chat endpoint (`/v1/chat/completions`).
+Then query it like any OpenAI chat endpoint (`POST /v1/chat/completions`).
+
+### What we learned (tips that matter)
+- **`--tool-call-parser qwen3_coder`** is required for clean OpenAI-style
+  `tool_calls`. Without it, tool calls come back as raw `<tool_call>` text and
+  break agent frameworks.
+- **`enable_thinking: false`** keeps internal reasoning out of the final answer.
+- **bf16 KV cache (`--kv-cache-dtype auto`)** for quality. FP8 KV roughly
+  doubles capacity but we observed quality drift on long agentic sessions, so
+  we keep bf16.
+- **Prefix caching (RadixCache) is the big win for agents.** In agentic loops
+  the same context is replayed every turn — cache-read hit rates of ~95%+ make
+  prefill almost free. Keep the agent context **append-only** to maximize hits.
+- Speculative decoding (NEXTN) did **not** help on this model under
+  concurrency — it lowered throughput, so we left it off.
+
+### Performance (measured on 1× RTX PRO 6000 Blackwell 96GB, bf16)
+| Metric | Value |
+|---|---|
+| Single-stream decode | **~84 tok/s** |
+| Under concurrency (several streams) | ~50–75 tok/s per stream |
+| KV cache pool | ~595k tokens |
+| Concurrent sessions @ ~60k ctx | ~10 |
+| Concurrent multi-agent pentest pipelines | ~2–3 |
+| Max context | up to 256k tokens |
+
+A single 96GB GPU comfortably serves a team's security copilots plus a few
+parallel pentest pipelines. For higher concurrency, load-balance across GPUs.
 
 ## ⚠️ Legal Notice / Disclaimer