rlx-mistral

Runner scaffold for Mistral 3+ / Ministral on RLX. A thin wrapper over rlx-llama32's Llama32Runner that validates the GGUF general.architecture tag (mistral3 / mistral4) and delegates decode.

Status: stub (PLAN.md M4). The underlying rlx-llama32 builder does not yet apply Mistral 3's per-layer sliding-window mask or the Mistral-specific RoPE base, so runs emit some tokens but will not match the upstream reference until those deltas land. Mistral 1/2 ship as plain llama — use rlx-llama32 directly for those.

Public API

use rlx_mistral::MistralRunner;

let mut runner = MistralRunner::builder()
    .weights("Ministral-3B.gguf")
    .max_seq(512)
    .packed_weights(true)      // low-memory packed-GGUF path
    .build()?;                 // errors unless arch ∈ {mistral3, mistral4}

let prompt_ids: Vec<u32> = /* tokenize */ vec![];
let out = runner.generate_packed(&prompt_ids, 64, |tok| print!("{tok} "))?;
# anyhow::Ok(())

MistralRunner also exposes config() -> &LlamaBaseConfig and inner() / inner_mut() to reach the full Llama32Runner API. cli_run(&args) re-checks the arch tag and forwards to the shared rlx-llama32 CLI (no dedicated binary is built yet). The crate re-exports Llama32ConfigSource, Llama32Runner, and Llama32RunnerBuilder.

How it fits

Inference and config parsing live in rlx-llama32 and rlx-llama-base; this crate is just the arch gate. Sibling GGUF-arch runners built the same way: rlx-phi (working), rlx-cohere, rlx-granite, rlx-glm, rlx-gpt-oss.