RWKV-7-G1-0.1B_BRIK

RWKV-7 « G1 » 0.1B — BRIK (int4)

RWKV-7 G1 0.1B (BlinkDL, Apache-2.0) converted to the BRIK format so it runs inside the browser on the visitor's GPU (WebGPU) — no inference server. Converted from a F16 GGUF.

The lightest file in the Brimkern catalogue, and the fastest to decode.

Try it in one click

👉 https://brimkern.com/chat?model=romainkh14/RWKV-7-G1-0.1B_BRIK

Why RWKV here

RWKV-7 is recurrent: a fixed-size state (~1 MB) replaces the KV cache, so memory does not grow with the conversation — the opposite of a transformer, whose cache swells with every token. On a light device that is the difference between a conversation that holds and one that eventually saturates.

Measured — and what it is not for

Chrome, Apple Silicon laptop, production build: 62.9 tok/s decoding (×3 over the pre-resident path). Replayable benches in scripts/e2e/.

It does not read documents. On our widget's document-Q&A bench (sdk-rag.mjs, 2 rounds × EN/FR) it scores 6/24: it serves the canonical refusal while the correct fact sheet is selected and in front of it, and when it does read, it copies the whole sheet out — forbidden figures included. What holds are greetings and out-of-scope refusals: its known strengths (conversational, constrained classification), not document reading.

Published as-is because a negative measurement is worth more than a claim: if you want a sub-150 MB Apache-2.0 model for grounded question answering, this size is not enough, whatever the file format. Use it for chat and classification; for document Q&A see G1a 0.4B (10/12) or LFM2.5-230M (12/12).

Format

A .brik is a self-describing container: architecture, tokenizer and configuration travel inside the file, weights are pre-quantized in the layout the kernels read, and every layer is one contiguous HTTP range. Specification: BRIK_FORMAT.md.

License

Weights under Apache-2.0 (RWKV-7 G1, BlinkDL). The Brimkern engine itself is MIT.