Monad Bullshitter

Pleias' Monad (56M parameters) reasoning on your own GPU: careful-looking thinking, mostly wrong answers. Open the thinking trace; that is the show.

Pleias' Monad has 56 million parameters, less than half of GPT-2, and it reasons before it answers: it parses your question, rates its own confidence, ticks off a verification checklist, and then tells you the capital of France is the capital of France. It got 7 of 30 one-line fact questions right. Open the thinking trace; that is the show. It runs on your own GPU, in this tab, with 8-bit weights. The default weights carry a training-free edit that stops most of the repetition loops the original falls into; it does not make Monad know more. Switch to the original to compare. The third choice is the loop-fixed Monad fine-tuned on 45 papers shared on Bluesky's Paper Skygest feed in early November 2026: ask it about one by title or topic and it gets about a third of questions right, and it loops more on long requests than the default.

backend … Weights are 65 MB, downloaded once and kept in your browser cache.

Source for Monad to answer from (optional, kept for every question until cleared)

Monad was trained to answer from sources it is handed (SYNTH's retrieval exercises). Paste text here and it is sent after each question, each paragraph in its own numbered <source_n> block as in SYNTH (up to ten); answers then quote it in <ref> tags, shown indented. Asked about a cat paragraph this way, both models got four of four questions right. Without it, both say a cat is not a mammal.

Settings: system prompt and sampling

System prompt

Monad was trained on single questions with no system turn. In dry runs, a system prompt that described Monad made it talk about itself in 30 to 90% of answers to unrelated questions, and plain style instructions changed little. Without one it already says it is Monad.

Append to each question

Instructions inside the question are the form Monad was trained on. " Answer in at most three sentences." cut the median answer from 148 words to 62, though now and then the instruction becomes the topic.

Start Monad's reply with

skip the thinking (the text then starts the answer)

Written into Monad's own turn, so it continues from it. Inside the thinking it steers the reasoning: planted "Cats are mammals ●." and both models reasoned to yes. With thinking skipped, Monad tends to fall into a multiple-choice format, and a planted "Yes," can be contradicted by the next words. "I’m unsure about the specific details" (a phrase from SYNTH's hedged answers) makes it decline and end the turn, on questions it would have got right as well as wrong ones.

Temperature

Top-p

Repetition penalty

Max new tokens

Show the thinking trace while it streams Stop when Monad starts repeating itself Send earlier turns with each question. Monad was trained on single questions; given an earlier exchange, it tends to answer the earlier question again.

Run on