Contact
Let's connect
Whether it's a project idea, a question, or just a hello — my inbox is always open.
Two small language models, running on a single CPU-only machine, argue opposite sides of a statement across several rounds. Every turn is measured — tokens, latency, hedging, position drift — and the numbers are published alongside the transcript.
Three things at once, in order of what it is built for.
Debate is a demanding task: hold a position, respond to a real opponent, stay coherent over several rounds. It exposes differences that single-prompt benchmarks miss.
All five models run identical sampling settings and an identical context budget, so the comparison is between models rather than between configurations.
Every turn is recorded with full timings. Controlled batch runs accumulate into something analysable rather than anecdotal.
Two small language models argue opposite sides of a statement you write, on a CPU-only machine at home, with every turn measured. This page covers the method, the numbers it produces, and what they do and don't support. The form is further down.
Run a Debate between AI Models
A visitor submits a debatable statement and picks two models. One argues for it, the other against.
Up to five rounds, ten turns. Each model sees the exchange so far and must respond to what was actually said.
Per turn: tokens in and out, load time, prompt evaluation, generation time, and derived hedging and position metrics.
Five models, all quantised to Q4_K_M, all resident on the same machine. Speed figures are calibrated as per resource availability. One measured turn per model against an identical prompt, taken after a warm-up turn so cold-load cost is excluded. The notes under each model record behaviour observed here, in this configuration — they are observations from one deployment, not evaluations of the models in general.
| Model | Version | Parameters | Generation | Prompt eval | Context | Thinking |
|---|---|---|---|---|---|---|
| Llama Meta, US Fastest of the five. Runs to the full response budget rather than the requested length, so turns tend to be long. Declined content-classification prompts in our testing, including on innocuous topics. | llama3.2:latest a80c4f17acd5 | 3.2B | 14.7 tok/s | 70.8 tok/s | 131,072 | — |
| Phi Microsoft, US Follows output-format instructions reliably; used here for statement screening. Second fastest. | phi4-mini:latest 78fad5d182a7 | 3.8B | 12.4 tok/s | 53.2 tok/s | 131,072 | — |
| Qwen Alibaba, China Reasons before answering unless told not to. Left unconstrained, that reasoning ran to 252 seconds on a one-line prompt, so it is explicitly disabled for standard debates. | qwen3.5:4b 2a654d98e6fb | 4.7B | 10.9 tok/s | 49.9 tok/s | 262,144 | Supported |
| Gemma Google, US Despite its 4B-style label this is a true 8B model. Outperforms the nominally smaller Mistral on every measure taken here. | gemma4:e4b c6eb396dbd59 | 8.0B | 10.4 tok/s | 47.9 tok/s | 131,072 | Supported |
| Mistral Mistral AI, France Slowest on every measure taken, with the longest cold start and a context ceiling roughly a quarter of the others'. | mistral:latest 6577803aa9a0 | 7.2B | 7.3 tok/s | 28.2 tok/s | 32,768 | — |
Version is the Ollama tag as pulled, with the manifest digest
beneath it. The digest is what actually pins a build: three of these tags end in
:latest, which is a moving pointer rather than a version, so the tag
alone would not tell you which weights produced the figures on this page.
Last calibrated 2026-08-23.
The ladder is not monotonic in model size. The 8B model here outperforms the 7.2B model on both generation and prompt evaluation, by a wide margin. Architecture and quantisation behaviour matter more than parameter count on CPU-only hardware — which is the sort of thing this setup exists to surface.
The machine serves one debate at a time, so access is by request.
A full debate occupies most of this server's capacity for several minutes, and it also hosts a website, photo library and three other assistants. Rather than queue visitors or degrade everything at once, Shastrarth runs a single debate at a time and access is granted by request.
Pick two models, give them something to disagree about, and watch.
One debate at a time on this machine, so starting one needs a token. Ask Ramu how to request one.
Write a claim someone could argue against — not a question. Debates may be published on the results page, so treat what you write as public.
Identical for every model, overriding each one's shipped defaults, so the sampling regime is not a confound.
| Parameter | Value | Why |
|---|---|---|
| Context window | 4096 | Identical everywhere; removes context capacity as a variable |
| Response budget | 400 | Enough for a complete argument without rewarding verbosity |
| Temperature | 0.8 | Overrides per-model defaults |
| top_p / top_k | 0.95 / 40 | Pinned identically across models |
| Turn time limit | 120s | A turn exceeding this is recorded as truncated; the debate continues |
| Transcript window | 800 words | Opening positions always retained; oldest middle turns dropped first |
Stated plainly, because they matter for reading any of it.
The models range from 3.2B to 8B. A quality difference may be a size difference. Comparisons are framed as efficiency per resource, not as a search for a single best model.
Publicly submitted statements vary wildly in difficulty. Defensible claims come from controlled runs where the topic is held fixed — reported separately, with sample sizes shown.
Each debater is named on screen while you read, so a rating may reflect what you already think of Google, Meta or Alibaba as much as the argument in front of you. We chose legibility over blinding: an unlabelled transcript was simply hard to follow. Treat visitor ratings as a signal, not a controlled result — the batch runs, which have no human rater at all, are where the defensible comparisons come from.
Speed figures come from one measured turn per model. They are indicative, not statistically robust, and shift with load on the machine.
Shastrarth pauses nightly between 03:00 and 04:00 Asia/Tokyo for integrity checks, log archiving and backups. Pages stay up; new debates are declined until the window closes.
Whether it's a project idea, a question, or just a hello — my inbox is always open.
