docs: simplify and minimalize README

[gondaliyashreyan1]

This commit completely overhauls the main README.md to align with the project's identity as a minimal, zero-jank AI runtime. Bloat, unnecessary tables, and references characterizing the project simply as a "llama.cpp fork" have been removed. Instead, the focus is squarely on its advanced native features (TurboQuant, MTP/NextN) pulled from the llama-turbo backend context, presented in a clean and compact format.

---

PR created automatically by Jules for task 110620075784319338 started by @gondaliyashreyan1

[google-labs-jules]

👋 Jules, reporting for duty! I'm here to lend a hand with this pull request.

When you start a review, I'll add a 👀 emoji to each comment to let you know I've read it. I'll focus on feedback directed at me and will do my best to stay out of conversations between you and other bots or reviewers to keep the noise down.

I'll push a commit with your requested changes shortly after. Please note there might be a delay between these steps, but rest assured I'm on the job!

For more direct control, you can switch me to Reactive Mode. When this mode is on, I will only act on comments where you specifically mention me with @jules. You can find this option in the Pull Request section of your global Jules UI settings. You can always switch back!

New to Jules? Learn more at jules.google/docs.

---

For security, I will only act on instructions from the user who triggered this task.

[gondaliyashreyan1]

Write a complete, polished, expansive README.md for the GitHub repository:

https://github.com/Anvil-LLM/anvil

IMPORTANT OUTPUT RULE:
Your entire response must be the README content inside ONE markdown code block. Do not write ANYTHING outside the code block. Do not add an introduction, explanation, apology, commentary, or closing message outside it.

The repository is called Anvil.

Anvil's tagline:
"Forge anything. A zero-jank, single-binary local AI runtime."

Anvil is a terminal-first local LLM runtime focused on making local inference extremely easy without sacrificing developer control.

Core positioning:

• Single native binary
• Zero runtime dependencies for the normal runtime experience
• No mandatory background daemon
• In-process inference
• Local-first
• Roughly 10 MB binary size
• Designed for both ordinary users and developers
• Built around a simple UX while exposing detailed runtime controls

IMPORTANT UX DETAIL:
Do NOT describe Anvil as simply having "magic defaults", "smart defaults", or silently guessed configuration.

The actual workflow is:

1. The user launches Anvil for the first time.
2. Anvil presents an interactive TUI configuration interface.
3. The USER explicitly chooses their configuration.
4. Anvil saves that configuration into a persistent JSON profile.
5. Future runs can simply use the saved configuration.
6. Developers can override individual settings with CLI flags whenever necessary.

Make this distinction clear throughout the README.

The TUI and CLI are complementary:

• TUI = easy initial configuration and persistent configuration management.
• JSON profiles = persistent configuration.
• CLI flags = direct control and per-run overrides.

Anvil should be presented as BOTH user-friendly and developer-friendly.

Known features from the current repository:

• Zero Overhead:
  • no hidden blob storage
  • no telemetry
  • no always-on daemon
  • runs only when invoked
• Maximum Speed:
  • TurboQuant KV-cache compression
  • speculative decoding
  • MTP support for Gemma 4
  • NextN support for Qwen 3.6
• Hardware-Adaptive:
  • detects CPU/GPU/RAM
  • configures appropriate backend/settings
• Model ecosystem:
  • pulls GGUF models from Hugging Face
  • pulls models from Ollama
  • built-in resumable downloads
• Persistent model profiles/settings
• Terminal-first interface
• Native local inference
• Single-binary distribution
• Linux NVIDIA/CUDA support with prebuilt runtime
• CUDA toolkit is bundled in the Linux CUDA distribution, so users only need the NVIDIA driver
• Universal installer script

Current quickstart:

curl -fsSL https://anvil-llm.github.io/anvil/install.sh | sh

[google-labs-jules]

Write a complete, polished, expansive README.md for the GitHub repository:

https://github.com/Anvil-LLM/anvil

IMPORTANT OUTPUT RULE:
Your entire response must be the README content inside ONE markdown code block. Do not write ANYTHING outside the code block. Do not add an introduction, explanation, apology, commentary, or closing message outside it.

The repository is called Anvil.

Anvil's tagline:
"Forge anything. A zero-jank, single-binary local AI runtime."

Anvil is a terminal-first local LLM runtime focused on making local inference extremely easy without sacrificing developer control.

Core positioning:

• Single native binary
• Zero runtime dependencies for the normal runtime experience
• No mandatory background daemon
• In-process inference
• Local-first
• Roughly 10 MB binary size
• Designed for both ordinary users and developers
• Built around a simple UX while exposing detailed runtime controls

IMPORTANT UX DETAIL:
Do NOT describe Anvil as simply having "magic defaults", "smart defaults", or silently guessed configuration.

The actual workflow is:

1. The user launches Anvil for the first time.
2. Anvil presents an interactive TUI configuration interface.
3. The USER explicitly chooses their configuration.
4. Anvil saves that configuration into a persistent JSON profile.
5. Future runs can simply use the saved configuration.
6. Developers can override individual settings with CLI flags whenever necessary.

Make this distinction clear throughout the README.

The TUI and CLI are complementary:

• TUI = easy initial configuration and persistent configuration management.
• JSON profiles = persistent configuration.
• CLI flags = direct control and per-run overrides.

Anvil should be presented as BOTH user-friendly and developer-friendly.

Known features from the current repository:

• Zero Overhead:
  • no hidden blob storage
  • no telemetry
  • no always-on daemon
  • runs only when invoked
• Maximum Speed:
  • TurboQuant KV-cache compression
  • speculative decoding
  • MTP support for Gemma 4
  • NextN support for Qwen 3.6
• Hardware-Adaptive:
  • detects CPU/GPU/RAM
  • configures appropriate backend/settings
• Model ecosystem:
  • pulls GGUF models from Hugging Face
  • pulls models from Ollama
  • built-in resumable downloads
• Persistent model profiles/settings
• Terminal-first interface
• Native local inference
• Single-binary distribution
• Linux NVIDIA/CUDA support with prebuilt runtime
• CUDA toolkit is bundled in the Linux CUDA distribution, so users only need the NVIDIA driver
• Universal installer script

Current quickstart:

curl -fsSL https://anvil-llm.github.io/anvil/install.sh | sh

# Anvil

> **Forge anything.**
> A zero-jank, single-binary local AI runtime. 

Anvil is a terminal-first local LLM engine that runs natively. It is designed to make local inference extremely easy for users without sacrificing developer control. 

One binary. Zero runtime dependencies. No mandatory background daemon. Just pure, in-process inference optimized for maximum throughput.

## Core Positioning

Anvil is built for both ordinary users and developers. It strikes a balance between user-friendly configuration and detailed runtime control:

- **Single Native Binary:** Roughly 10 MB in size, distributed as a single file.
- **Zero Overhead:** No hidden blob storage, no telemetry, no always-on daemons. Anvil runs only when you invoke it.
- **In-Process Inference:** Local-first, native execution without unnecessary client-server abstraction.
- **Hardware-Adaptive:** Automatically detects your CPU, GPU, and RAM to help configure appropriate backend settings.

## How Configuration Works

Anvil does **not** rely on silent "magic defaults" that guess what you want. Instead, it uses a clear, explicit workflow that combines interactive setup with persistent configuration:

1.  **Initial Setup:** When you launch Anvil for the first time, it presents an interactive TUI (Terminal User Interface).
2.  **Explicit Choice:** You, the user, explicitly choose your desired configuration.
3.  **Persistent Profiles:** Anvil saves your choices into a persistent JSON profile.
4.  **Effortless Execution:** Future runs automatically use your saved configuration.
5.  **Developer Control:** You can always override individual settings on a per-run basis using CLI flags.

**The Anvil Workflow:**
*   **TUI** = Easy initial configuration and persistent profile management.
*   **JSON Profiles** = Persistent, transparent configuration storage.
*   **CLI Flags** = Direct control and immediate, per-run overrides.

## Maximum Speed Features

Anvil leverages advanced techniques to extract the maximum performance from your hardware:

- **TurboQuant:** Built-in KV-cache compression.
- **Speculative Decoding:** Supports advanced speculative decoding methods to accelerate generation.
- **MTP Support:** Integrated support for Gemma 4 Multi-Token Prediction.
- **NextN Support:** Integrated support for Qwen 3.6 NextN speculative decoding.

## Seamless Ecosystem Integration

Anvil plays nicely with the rest of the open-weight ecosystem:

- **Hugging Face:** Pull GGUF models directly from Hugging Face repositories.
- **Ollama:** Pull models directly from the Ollama registry.
- **Reliable Networking:** Built-in resumable downloads ensure large models arrive intact.

## Quickstart

**Install (Universal Installer):**
```bash
curl -fsSL https://anvil-llm.github.io/anvil/install.sh | sh

(On Linux x86_64, the installer automatically prefers the CUDA prebuilt which bundles the CUDA toolkit — you only need the NVIDIA driver).

Run a local model:

anvil run model.gguf

Pull and run from Ollama:

anvil pull ollama:llama3.2:3b
anvil run llama3.2:3b

Interactive pull from HuggingFace:

anvil pull hf:bartowski/Llama-3.2-1B-Instruct-GGUF

Advanced Usage

Once configured via the TUI, you can still use the CLI for granular control.

Override settings for a single run (maximize performance):

anvil run model.gguf --ctx 128000 --type-k turbo4 --type-v turbo3 --mtp

Save overrides to the model's persistent profile:

anvil run llama3.2 --temp 0.9 --save

Manage profiles directly:

anvil profile llama3.2 set n_ctx=131072 type_k=turbo4

NVIDIA / CUDA Support (Linux)

The Linux CUDA distribution of Anvil bundles the CUDA toolkit. This means you do not need to install CUDA yourself; you only need the official NVIDIA driver.

If you are missing drivers, you can use our universal installer script which supports any distro and GPU:

curl -fsSL -O https://raw.githubusercontent.com/anvil-llm/anvil/main/docs/anvil-nvidia-install.sh
sudo sh anvil-nvidia-install.sh

Build from Source

git clone --recursive https://github.com/anvil-llm/anvil
cd anvil
cmake -B build -S . -DCMAKE_BUILD_TYPE=Release
cmake --build build -j

License

MIT