I hid an entire webpage inside a cat face

gen — Mon, 18 May 2026 13:46:38 +0000

The source of this page is just a cat face:

https://nothing-to-see-here.surge.sh/

View source. You'll see (=･ω･=) and nothing else.
But the page runs. Rainbow animation, layout, everything.

The entire JavaScript is encoded as invisible Unicode
Variation Selectors attached to the cat emoji.

Unicode VS (U+FE00–FE0F and U+E0100–E01EF) map
precisely to 256 byte values. Any byte sequence can
ride inside normal text, invisible to readers,
surviving copy-paste across Slack, X, LINE, iMessage.

That's subtext.

https://prospectorlabs.dev/subtext

267 tok/s local inference on RTX 5090 – llama.cpp MTP + Qwen3-35B-A3B MoE

gen — Mon, 18 May 2026 12:56:14 +0000

Been running Qwen3-35B-A3B (MoE) with llama.cpp's Multi-Token Prediction
(MTP / speculative decoding) on an RTX 5090 under WSL2. Results surprised me:

Model	Speed
Ollama stock (35B MoE)	171 tok/s
27B Dense + MTP	104 tok/s
35B MoE + MTP	267 tok/s ← this

For context: Claude Haiku runs ~150 tok/s via API, billed at $150/MTok.
This setup runs on electricity only.

The interesting finding is that MoE and speculative decoding have unusual
synergy. With a dense model, MTP gave a modest speedup (or none).
With MoE, it nearly doubled throughput.

My hypothesis: MoE's sparse activation pattern leaves compute headroom that
speculative decoding can exploit. The draft tokens are cheap to verify because
most experts stay inactive during verification passes.

Setup:

RTX 5090, WSL2 (Ubuntu 24)
llama.cpp with MTP draft, n-max 2
Qwen3-35B-A3B-Instruct Q4_K_XL
ctx 65536, OpenAI-compatible API on localhost

Happy to share the exact llama-server launch flags if anyone wants to reproduce.

DEV Community: gen

I hid an entire webpage inside a cat face

267 tok/s local inference on RTX 5090 – llama.cpp MTP + Qwen3-35B-A3B MoE