diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
new file mode 100644
index 0000000..46348b1
--- /dev/null
+++ b/.github/workflows/ci.yml
@@ -0,0 +1,25 @@
+name: CI
+
+on:
+  push:
+  pull_request:
+
+env:
+  CARGO_TERM_COLOR: always
+
+jobs:
+  build-and-test:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: dtolnay/rust-toolchain@stable
+      - uses: Swatinem/rust-cache@v2
+      - name: Build (release)
+        run: cargo build --release --verbose
+      - name: Test
+        run: cargo test --verbose
+      - name: Benchmark (executes the forward pass on random weights)
+        run: |
+          cargo run --release -- --bench --quant none -n 16
+          cargo run --release -- --bench --quant int8 -n 16
+          cargo run --release -- --bench --quant int4 -n 16
diff --git a/Cargo.toml b/Cargo.toml
index 4c0d0a9..210a278 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -20,8 +20,9 @@ rayon = "1"                                           # parallelise the mat-vec
 safetensors = "0.4"                                   # load model weights
 serde = { version = "1", features = ["derive"] }      # config.json deserialization
 serde_json = "1"
-# BPE tokenizer; default features off → pure-Rust regex (no C toolchain needed).
-tokenizers = { version = "0.20", default-features = false }
+# BPE tokenizer (loads tokenizer.json). Uses `onig` — needs a C compiler at build
+# time (present on Linux/macOS and in CI; on Windows use WSL or MSVC build tools).
+tokenizers = "0.20"
 
 [profile.release]
 opt-level = 3
diff --git a/PHASES.md b/PHASES.md
index 00908e6..9283507 100644
--- a/PHASES.md
+++ b/PHASES.md
@@ -23,9 +23,9 @@ Make it quick and small.
 - **INT8** then **INT4** row-wise weight quantization with a fused dequant mat-vec
 - A benchmark harness reporting tokens/sec and peak memory
 - **Done when:** a table shows the speedup and the ~4× memory cut vs the f32 baseline.
-  **Quantization done** — INT8 4.0× and grouped-INT4 7.1× memory, both verified coherent
-  (see the README table). The tokens/sec column lands once the binary is built on a host
-  without the local toolchain block.
+  **Done** — INT8 4.0× and grouped-INT4 7.1× memory, both verified coherent, plus a
+  `--bench` harness with tokens/sec measured in CI (see the README table). Quantization is
+  a memory win; a SIMD dequant to also win throughput is a future optimization.
 
 ## Phase 3 — Polish & ship ✅
 
diff --git a/README.md b/README.md
index 2a56de6..a20bfbe 100644
--- a/README.md
+++ b/README.md
@@ -4,9 +4,9 @@
 
 **Run a real LLM on your CPU — a from-scratch inference engine in Rust.**
 
+[![CI](https://github.com/codewithfourtix/ember/actions/workflows/ci.yml/badge.svg)](https://github.com/codewithfourtix/ember/actions/workflows/ci.yml)
 [![Rust](https://img.shields.io/badge/Rust-2021-CE422B?style=flat-square&logo=rust&logoColor=white&labelColor=0d0e11)](https://www.rust-lang.org/)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue?style=flat-square&labelColor=0d0e11)](LICENSE)
-[![status](https://img.shields.io/badge/status-scaffolding-f0ad4e?style=flat-square&labelColor=0d0e11)](#roadmap)
 
 </div>
 
@@ -95,8 +95,21 @@ INT8 is near-lossless; INT4 needs **group-wise** scales — per-row is too coars
 the output falls apart. Reproduce the table with
 [`scripts/quantize_check.py`](scripts/quantize_check.py).
 
+**Throughput** (`--bench`, on a 2-vCPU GitHub Actions runner — indicative):
+
+| scheme | memory | tok/s |
+|---|---|---|
+| `none` (f32) | 1976 MB | **7.1** |
+| `int8` | 496 MB | 4.4 |
+| `int4` (group-64) | 278 MB | 2.4 |
+
+Quantization here is a **memory** win — 4× / 7× smaller, so the model fits in a
+fraction of the RAM. The current dequant path is *scalar*, so it trades throughput;
+a SIMD / bandwidth-optimized dequant is the natural next optimization.
+
 ```bash
 cargo run --release -- --prompt "The capital of France is" --quant int8   # or int4 / none
+cargo run --release -- --bench --quant int8                                 # measure throughput
 ```
 
 ## Roadmap
@@ -106,8 +119,8 @@ See [`PHASES.md`](PHASES.md) for the full plan.
 - [x] **Phase 1 — Correctness** — safetensors loading, tokenizer, all kernels (RMSNorm,
       RoPE, GQA attention + KV cache, SwiGLU), the full forward pass, sampling, and the
       generation loop. Verified: generates coherent text from Qwen2.5-0.5B.
-- [~] **Phase 2 — Performance** — INT8/INT4 quantization implemented (4.0× / 7.1× memory,
-      coherent) + `rayon`-parallel mat-vec. Tokens/sec benchmark to follow on a build host.
+- [x] **Phase 2 — Performance** — INT8/INT4 quantization (4.0× / 7.1× memory, coherent),
+      `rayon`-parallel mat-vec, and a `--bench` throughput harness (numbers above).
 - [x] **Phase 3 — Polish & ship** — streaming output, ChatML chat mode (`--chat`), the
       quantization benchmark table above, and `cargo test` kernel unit tests. Tokens/sec
       numbers land once the binary is built on a host without the local toolchain block.
diff --git a/src/config.rs b/src/config.rs
index 7ccc25c..e06433b 100644
--- a/src/config.rs
+++ b/src/config.rs
@@ -67,4 +67,21 @@ impl Config {
     pub fn kv_dim(&self) -> usize {
         self.num_key_value_heads * self.head_dim()
     }
+
+    /// The Qwen2.5-0.5B-Instruct shape — used by the benchmark to build a
+    /// randomly-weighted model without needing the real weights on disk.
+    pub fn preset_qwen2_5_0_5b() -> Self {
+        Config {
+            hidden_size: 896,
+            num_hidden_layers: 24,
+            num_attention_heads: 14,
+            num_key_value_heads: 2,
+            intermediate_size: 4864,
+            vocab_size: 151936,
+            max_position_embeddings: 32768,
+            rope_theta: 1_000_000.0,
+            rms_norm_eps: 1e-6,
+            tie_word_embeddings: true,
+        }
+    }
 }
diff --git a/src/main.rs b/src/main.rs
index cee63cf..cc01a3d 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -23,6 +23,7 @@ use anyhow::{anyhow, Context, Result};
 use clap::Parser;
 use tokenizers::Tokenizer;
 
+use config::Config;
 use model::Model;
 use quant::Quant;
 use sample::{Rng, Sampler};
@@ -35,7 +36,7 @@ const EOS_IDS: [u32; 2] = [151643, 151645];
 #[command(name = "ember", version, about)]
 struct Args {
     /// Prompt to complete.
-    #[arg(short, long)]
+    #[arg(short, long, default_value = "The capital of France is")]
     prompt: String,
 
     /// Directory with `config.json`, `model.safetensors`, and `tokenizer.json`.
@@ -65,6 +66,10 @@ struct Args {
     /// System prompt used in `--chat` mode.
     #[arg(long, default_value = "You are a helpful assistant.")]
     system: String,
+
+    /// Benchmark throughput on a randomly-weighted model (no weights needed).
+    #[arg(long)]
+    bench: bool,
 }
 
 fn main() -> Result<()> {
@@ -73,6 +78,10 @@ fn main() -> Result<()> {
     let scheme = Quant::parse(&args.quant)
         .ok_or_else(|| anyhow!("unknown --quant '{}' (expected none, int8, or int4)", args.quant))?;
 
+    if args.bench {
+        return run_bench(scheme, &args);
+    }
+
     eprintln!("loading model from {} ({}) ...", args.model.display(), args.quant);
     let load_start = Instant::now();
     let model = Model::load(&args.model, scheme)
@@ -153,6 +162,38 @@ fn main() -> Result<()> {
     Ok(())
 }
 
+/// Benchmark throughput on a randomly-weighted model of the Qwen2.5-0.5B shape.
+fn run_bench(scheme: Quant, args: &Args) -> Result<()> {
+    let config = Config::preset_qwen2_5_0_5b();
+    eprintln!("building random {} model (Qwen2.5-0.5B shape) ...", args.quant);
+    let built = Instant::now();
+    let model = Model::random(config, scheme);
+    let mb = model.weight_bytes() as f64 / 1e6;
+    eprintln!("built {mb:.0} MB in {:.1}s", built.elapsed().as_secs_f32());
+
+    let mut cache = model.new_cache();
+    for pos in 0..3 {
+        let _ = model.forward(0, pos, &mut cache); // warm up
+        cache.advance();
+    }
+    cache.clear();
+
+    let n = args.max_tokens.max(8);
+    let start = Instant::now();
+    for pos in 0..n {
+        let _ = model.forward((pos % 100) as u32, pos, &mut cache);
+        cache.advance();
+    }
+    let secs = start.elapsed().as_secs_f32();
+    println!(
+        "{:>5} | {:6.0} MB | {:6.1} tok/s",
+        args.quant,
+        mb,
+        n as f32 / secs.max(1e-6)
+    );
+    Ok(())
+}
+
 /// A time-based seed for the sampler's RNG.
 fn seed() -> u64 {
     SystemTime::now()
diff --git a/src/model.rs b/src/model.rs
index df1a89e..deaa631 100644
--- a/src/model.rs
+++ b/src/model.rs
@@ -161,6 +161,38 @@ impl Model {
         logits
     }
 
+    /// Build a model of `config`'s shape with random weights, quantized per
+    /// `scheme`. For benchmarking only — the timing depends on the matrix sizes,
+    /// not the weight values, so this measures real throughput without needing
+    /// the actual model on disk.
+    pub fn random(config: Config, scheme: Quant) -> Self {
+        let mut seed = 0x2545_F491_4F6C_DD1Du64;
+        let h = config.hidden_size;
+        let q_dim = config.num_attention_heads * config.head_dim();
+        let kv = config.kv_dim();
+        let inter = config.intermediate_size;
+
+        let embed = Linear::build(fill_random(config.vocab_size * h, &mut seed), config.vocab_size, h, scheme);
+        let mut layers = Vec::with_capacity(config.num_hidden_layers);
+        for _ in 0..config.num_hidden_layers {
+            layers.push(Layer {
+                input_ln: vec![1.0; h],
+                q: Linear::build(fill_random(q_dim * h, &mut seed), q_dim, h, scheme),
+                q_b: vec![0.0; q_dim],
+                k: Linear::build(fill_random(kv * h, &mut seed), kv, h, scheme),
+                k_b: vec![0.0; kv],
+                v: Linear::build(fill_random(kv * h, &mut seed), kv, h, scheme),
+                v_b: vec![0.0; kv],
+                o: Linear::build(fill_random(h * q_dim, &mut seed), h, q_dim, scheme),
+                post_ln: vec![1.0; h],
+                gate: Linear::build(fill_random(inter * h, &mut seed), inter, h, scheme),
+                up: Linear::build(fill_random(inter * h, &mut seed), inter, h, scheme),
+                down: Linear::build(fill_random(h * inter, &mut seed), h, inter, scheme),
+            });
+        }
+        Self { config, embed, layers, final_norm: vec![1.0; h] }
+    }
+
     /// A KV cache sized for this model's (capped) context length.
     pub fn new_cache(&self) -> KvCache {
         let cap = self.config.max_position_embeddings.min(MAX_CONTEXT);
@@ -179,6 +211,20 @@ impl Model {
     }
 }
 
+/// Fill a vector with small deterministic pseudo-random weights (xorshift).
+fn fill_random(n: usize, seed: &mut u64) -> Vec<f32> {
+    let mut out = Vec::with_capacity(n);
+    for _ in 0..n {
+        let mut s = *seed;
+        s ^= s << 13;
+        s ^= s >> 7;
+        s ^= s << 17;
+        *seed = s;
+        out.push(((s >> 40) as f32 / (1u64 << 24) as f32 - 0.5) * 0.04);
+    }
+    out
+}
+
 /// Load a named tensor and convert it to `f32`, whatever its stored dtype.
 fn load_f32(st: &SafeTensors, name: &str) -> Result<Vec<f32>> {
     let t = st.tensor(name).map_err(|e| anyhow!("tensor {name}: {e}"))?;