Blog: t27b v0.2.0 -- a 3 MB compiler that turns a spec into AArch64 code in 5 ms, and where it still loses #1600

Problem

t27b v0.2.0 was released on 2026-10-10 (https://github.com/gHashTag/t27/releases/tag/t27b-v0.2.0) with the signed receipt of the binary it ships, and its benchmark against zig, gcc, clang and rustc is written up in gHashTag/t27#8855 (issue gHashTag/t27#8817). t27.ai has no post about either.

Plan

One post, a-3-mb-compiler-spec-to-aarch64-in-5-ms, English and Russian:

• the hook: a compiler of about 3 MB that turns a spec into AArch64 code in about 5 ms;
• the method and its caveats: one shared lab, qemu-emulated run times, a spec set chosen without looking at t27b;
• four figures drawn from the report's JSON and the t27b lab's run summaries: compile time on a log scale, generated code against gcc -O0/-O2 and zig, memory and footprint, and the week's coverage climb;
• two tables: the per-axis summary, and where t27b loses;
• the trust section (signed receipts, the JIT/interpreter check, the limits of DDC) and what comes next;
• a new terminal recording, made on the Railway t27c lab, in both bodies;
• its own og-art triptych in gHashTag/ghashtag.github.io.

Every number comes from the report or from the lab's signed receipt for t27 59e2c87c4, the release commit.