Skip to content
Torad. LAB / 001
Talk to us

Research

Internal model measurements, published technical notes, and methods that are not public yet.

Measurements ↓Research blog →

ScopeEvery result names its model and comparison. None has been independently reproduced.

Measurements

Five readings from current Torad runs.

Output similarity
0.9998

Qwen3-1.7B in TQ4. The top answer matched on five of five prompts.

Reference: 1.0000 bf16 cosine
Linear weights870 MB

Qwen3-1.7B in TQ4.

Reference: 2.7 GB bf16
TQ4 runtime238 tok/s

Qwen3-1.7B.

Reference: 267 tok/s on vLLM
Native FP440 / 21

Qwen3-8B and Qwen3-14B on 16 GB.

Tokens per second
Training peak13.5 GB

Gemma 4 E2B on an RTX 5080.

16 GB available

Published notes

The blog carries experiment history, reasoning, and later corrections.

  1. 01SFT copies, GRPO discoversTraining

    Why imitation and reinforcement produce different kinds of learning.

  2. 02Rejection architectureMethod

    Refusal as part of the system design.

  3. 03Impedance matching in communicationTheory

    Transfer between structurally matched systems.

  4. 04Manufactured discoveryMethod

    Training a model to find an answer.

Methods still being evaluated

These artifacts exist internally. Their full methods or evaluations are not public.

Trainable 4-bit controls
The controls above the frozen 4-bit substrate have not been published.
Megakernel
The code-size artifact is recorded; the complete method and evaluation are not public.
QGRE
An internal Hamiltonian derivation run exists; the method and token-level evaluation are not public.