README polish: TL;DR up top, inline chat template

Two small documentation improvements:

1. Move the TL;DR (3-line 'just install and run' block) to right after
the intro paragraph, before 'Why a 27B variant'. Someone landing on
the model card now sees a working command before any rationale.

2. Inline the chat-template section instead of pointing readers at the
35B sibling's README. Three short examples: plain conversation,
reasoning trace with <think>, and an XML <tool_call>. README now
stands on its own without a clickthrough.

3. Drop the 'Same as the 35B sibling:' lead-in above the system prompt
(the prompt was already inlined; the lead-in was just clutter).

CHANGELOG.md

@@ -7,6 +7,19 @@ and documentation**, not the underlying base model.
 
 ## [Unreleased]
 
+### Changed
+- README: added a TL;DR section right after the intro paragraph so
+  someone scanning the page gets a working command without scrolling
+  past Why / What's here / Architecture.
+- README chat-template section: replaced the cross-reference to the
+  35B sibling card with inlined examples (plain conversation,
+  `<think>` reasoning trace, XML tool call). The README now stands
+  on its own.
+- Minor: dropped the "Same as the 35B sibling:" lead-in above the
+  system prompt block; the prompt was already inlined.
+
+## [0.5.0] - 2026-05-02 — `e4beea4`
+
 ### Added
 - `examples/llama_cpp_vision.py` — image-text-to-text via
   `llama-cpp-python` + a separate `mmproj-F16.gguf`. Currently the only

README.md

@@ -58,6 +58,27 @@ pipeline_tag: image-text-to-text
 
 A personal sibling to [`FoolDev/janus`](https://huggingface.co/FoolDev/janus). Same teacher (Claude Opus 4.7), same dataset family, but built on the **dense** [Qwen/Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B) base instead of the 35B-A3B MoE. Smaller, easier to deploy, no expert-routing surprises.
 
+## TL;DR
+
+If you have Ollama and 24 GB of RAM (or a 24 GB GPU):
+
+```bash
+git clone https://huggingface.co/FoolDev/janus-27b && cd janus-27b
+make build                            # downloads ~17 GB GGUF and creates the model
+ollama run janus-27b
+```
+
+If you're on a 32 GB unified-memory laptop (Mac M-series, Z13, etc.) use
+the smaller profile:
+
+```bash
+make build PROFILE=z13 QUANT=Q3_K_S   # ~12 GB GGUF, fits in ~17 GB total
+ollama run janus-27b-z13
+```
+
+For image input use llama.cpp directly — Ollama vision is broken for
+this architecture upstream (see [Vision](#vision)).
+
 ## Why a 27B variant?
 
 The 35B-A3B is a sparse mixture-of-experts model: 35B parameters total but only ~3B active per token. That makes it fast at inference but **memory-hungry at load time** — the full 35B has to live in VRAM/RAM even though only 3B is doing useful work each step.
@@ -167,7 +188,8 @@ Lower temperature (0.4-0.6) and bump `repeat_penalty` to 1.08 if it loops inside
 
 ### System prompt
 
-Same as the 35B sibling:
+The Modelfile bakes this in. Override per-request via the `system` role
+in your client:
 
 ```text
 You are Janus, a precise and capable assistant for reasoning, writing, coding, and long-form dialogue.
@@ -240,9 +262,56 @@ The dense 27B is the easier of the two Janus models to deploy.
 
 ## Chat template
 
-Identical to the 35B sibling — Qwen 3.x ChatML with `<|im_start|>` / `<|im_end|>` markers, `<think>...</think>` for reasoning traces, XML-style `<tool_call>` for function calling. The template is embedded in the GGUF metadata.
+Standard Qwen 3.x ChatML with `<|im_start|>` / `<|im_end|>` role markers,
+`<think>...</think>` blocks for reasoning traces, and XML-style
+`<tool_call>` for function calling. The template is embedded in the GGUF
+metadata, so loaders that read it (llama.cpp, Ollama, LM Studio) handle
+the formatting automatically.
+
+#### Plain conversation
+
+```text
+<|im_start|>system
+You are Janus, a precise and capable assistant…<|im_end|>
+<|im_start|>user
+What is the time complexity of mergesort?<|im_end|>
+<|im_start|>assistant
+```
+
+#### With reasoning trace
+
+```text
+<|im_start|>assistant
+<think>
+The user asked about mergesort. It splits, recursively sorts each half,
+then merges. The recurrence T(n) = 2T(n/2) + O(n) solves to O(n log n).
+</think>
+
+Mergesort runs in **O(n log n)** time in the worst, average, and best
+cases.<|im_end|>
+```
+
+Most clients (Open WebUI, LibreChat, etc.) hide the `<think>` block by
+default and surface only the visible answer. Strip it manually with
+`re.sub(r"<think>.*?</think>\s*", "", content, flags=re.DOTALL)` if your
+client doesn't.
+
+#### Tool / function calling
+
+The embedded template uses Qwen's XML format:
+
+```text
+<tool_call>
+<function=get_current_weather>
+<parameter=city>Paris</parameter>
+<parameter=unit>celsius</parameter>
+</function>
+</tool_call>
+```
 
-See the [Janus-35B Chat template section](https://huggingface.co/FoolDev/janus#chat-template) for examples — they apply unchanged here.
+Most OpenAI-compatible servers (Ollama, LM Studio, vLLM) translate
+between this and the JSON `tool_calls` shape automatically. See
+`examples/ollama_chat.py:tool_round_trip` for a working round-trip.
 
 ## Known limitations