Drop findings examples; let the link carry it

README.md

@@ -17,11 +17,7 @@ The model shelf of [**protoLabs.studio**](https://protolabs.studio): an indie st
 
 We run 35B-class open models on a pair of prosumer Blackwell GPUs, and the builds that make that possible go up here for anyone to pull: **FP8 quants** and **speculative-decode heads**, parity-verified against the source before we ship. The recipe and the verification are the product, not just the weights.
 
-The findings behind them get written up at [protolabs.studio](https://protolabs.studio):
-
-- CUDA graphs run **37–470% faster** on Blackwell
-- INT4 holds quality on dense models but corrupts MoE routing — so MoE stays bf16
-- `NCCL_P2P_DISABLE=1` clears TP=2 corruption on PCIe: **9.3×** on a 35B MoE
+The findings behind them get written up at [protolabs.studio](https://protolabs.studio).
 
 ## The rest of the studio