v0.3.0 - Kimi K3 + DeepSeek-V4-Flash-0731

Both architectures supported and verified 1:1 against the official reference code, unified CLI (--arch k3|dsv4), Metal GPU on macOS. Assets: nanokimi-0.2b.bin + vocab_nano.json (the pretrained test model, TinyStories). The debug models (microkimi-debug.bin, microdeepseek-debug.bin) are NOT shipped - generate them locally with: microkimi build / microkimi build --arch dsv4

v0.2.0 - Metal GPU support (macOS) + nanokimi pretrained model

Adds zero-dependency Metal GPU support on macOS (--gpu, weights cached on device, tuned for large matvecs), plus a leaner README. Includes nanokimi-0.2b.bin + vocab_nano.json: the pretrained 0.2B model (TinyStories, overnight CPU training, no GPU). Download both files into the repo root, then: cargo build --release && ./target/release/microkimi run "Once upon a time" --model nanokimi-0.2b.bin --max-new 12 --raw

v0.1.0 - nanokimi pretrained model

Pretrained nanokimi-0.2b (181.6M params, TinyStories, overnight CPU training, no GPU). Download both files into the repo root, then: cargo build --release && ./target/release/microkimi run "Once upon a time" --model nanokimi-0.2b.bin --max-new 12 --raw