From eb6e00759cc5162cedda4a992f160999e782d848 Mon Sep 17 00:00:00 2001 From: julianmb Date: Tue, 1 Sep 2026 10:23:36 +1000 Subject: [PATCH] readme: add plain-english collapse at top --- README.md | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+) diff --git a/README.md b/README.md index 2f0555c..21b7897 100644 --- a/README.md +++ b/README.md @@ -16,6 +16,27 @@ +
+plain english — what was done + +qwen3.8-flash-next is the new qwen model that is great for coding — it beats +claude opus 4.6 on swe-bench and runs on a $2500 mini pc. + +1. the official conversion code missed a step — every hyper-connection norm + was off by exactly 1.0, so the first quant printed garbage. we found it, + fixed the converter, and added a test so it never happens again. +2. we made the model smaller without losing quality — the big 51b n-gram table + tolerates 4-bit, saving 27g — then proved it across context depths from + 0 to 256k. +3. we kept everything that makes strix halo fast — vulkan kernels, graph + reuse, speculative decoding with the 4b draft head. + +if you just want to run it: `docker compose up --build` and open +`http://localhost:8080`. pick the 91g file for speed, the 116g file if you +need 262k context. + +
+ --- ## results