Timothy. Z @TimothyZ77

Our house in Toronto had this fancy glassed wine cellar, but we are not big wine drinkers. Anyways, it had good climate control so here is my rack cellar.

Median controller latency: Qwen 419 ms, Jev 135 ms. Code supplies structured board data, legal moves and projected outcomes. One short run make it self-explanatory Code and traces: github.com/TimothyZhang7/…

I added a Tetris example to Open Decisions, using local Qwen3-VL-4B alongside hosted Jev. Qwen chooses actions from logits, with no generated answer text. The video shows one 60-second run per backend, starting with the same board and piece sequence.

Timothy. Z @TimothyZ77

3 days ago

Glad Jev is bringing attention to a pattern we've been looking for: fast, typed decisions. I think wider adoption could make agents much more efficient. I built Open Decisions to explore this locally, with zero generated answer tokens: github.com/TimothyZhang7/…

2 0 1 10K 0

@dakshgup But the hammer is cheap!

@inco_ai sheesh

Qwen3.8-27B at 144 tok/s on an M5 Max MacBook Pro ⚡ Meet Inco Splash: our open-source inference engine, built around the model and around Apple silicon. Up to 3× the decode speed of Ollama, 2× oMLX, and almost 4× when an agent fans out into sub-agents.

@kunchenguid Simple, real human don’t speak with 10 consecutive punchlines

@LukeberryPi Text

@thdxr Do you believe I’m a bot?

@thdxr Why not

@trq212 finally

@jjacky That’s only 10 tps unfortunately

@elder_plinius tldr: LLM that produce no text, only probability distributions and convert to JSON

@CompleteSkeptic Jev prompted me to explore this with open models. I built a small Python package for typed decisions using local Qwen, and tried it on Tetris alongside Jev. Code, traces and limitations are public. github.com/TimothyZhang7/…

The clip shows 113–533 ms engine times on an M4 Pro, with the model already loaded. Cache status is shown. The probabilities are uncalibrated. The repo includes evaluations and a case where the model confidently reversed two images.

Glad Jev is bringing attention to a pattern we've been looking for: fast, typed decisions. I think wider adoption could make agents much more efficient. I built Open Decisions to explore this locally, with zero generated answer tokens: github.com/TimothyZhang7/…

The demo uses Qwen3-VL-4B. It reads the logits for allowed answers and returns two probabilities: whether to delegate and whether to ask for context. Both come from one model evaluation. The model still runs inference; it generates no answer tokens.

@MaxRovensky That's funny