KLM 1.6

KLM 1.6

Continues English text with a tiny 99M base language model from MinimaLabs. For researchers and hobbyists who want a small model to study, test or fine-tune.

Made by MinimaLabs on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: KLM 1.6 is a base language model with 99M parameters from MinimaLabs. It carries on training from a KLM 1.5 checkpoint saved before that model's decay stage. The main change from 1.5 is a longer decay of the learning rate.

Who it suits: It suits researchers and hobbyists who want a tiny English model to study, compare or build on.

Industry | Who uses it | How they use it | Example
Research | Small-model researchers | Compare tiny models on shared benchmarks | A check against SmolLM2-135M on ARC and PIQA
Software | Developers | Try a text model with a tiny memory footprint | A prototype that loads a 99M model locally
Education | Students | Study how a small model is trained in stages | A class project on warm-up and decay

People who want a chat assistant can skip it, since this is a base model and not one tuned to follow instructions.

Where to use it:

• Hugging Face Transformers, the library the card lists
• Windows, macOS or Linux, where you download the files with Python
• Machines with little memory, since the weights are about 396 MB in F32 or about 49 MB at 4-bit by estimate

Use case | Who | What they get back
Text continuation | Developers | English text that carries on from your prompt
Benchmark comparison | Researchers | Scores on ARC, HellaSwag, PIQA and ArithMark-3
Fine-tuning base | Hobbyists | A small starting point for your own training

How to set it up: You need these first:

• Python and the huggingface_hub package
• The transformers library, since the card lists it
• A computer with a few hundred MB of free memory

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Load the model with transformers.

pip install -U huggingface_hub
hf download MinimaLabs/KLM-1.6

How to use it: Load the model in transformers and give it the start of a sentence or paragraph in English. A base model carries the text on rather than answering like an assistant, so write prompts as text to be continued.

Example: You give it the opening of a short paragraph on a simple topic. It writes a few more sentences in the same style. A model this small can drift or repeat, so treat the output as a test, not a finished text.

For newcomers and experts: A base model has learned from lots of text to predict what comes next. A first small thing to try is giving it half a sentence and reading how it finishes.

For experts, the model has 24 layers that mix Gated DeltaNet-2 and attention in a 3 to 1 ratio, a hidden size of 512 and a context of 2,048 tokens. It also holds an n-gram memory of 125,000 bigrams and 38,000 trigrams, and its vocabulary holds only 4,100 tokens. Training ran in three stages, a superposition warm-up, a stable stage and a decay. On the Open SLM leaderboard the card reports an Intelligence Index of 33.95, first among models under 100M, 150M and 250M parameters.

Warm up with token superposition training
Train steadily at a constant learning rate
Decay the learning rate over 100B tokens
Load the result and continue your text

Good to know:

Title: ArithMark-3 score
KLM 1.6: 55.90%
GPT-X3-150M: 50.90%
KLM 1.5: 48.20%
SmolLM2-135M: 39.20%

• Size: 99M parameters in Safetensors format
• Language: English
• Context: 2,048 tokens
• License: apache-2.0
• Cost to run: none named, no API key or account
• Data it reaches: it downloads model files from Hugging Face and works on the text you give it
• Not for: none named in the card text shown

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 99M
• Weights: about 396 MB in F32; about 49 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 909
• Likes: 60
• License: apache-2.0
• Languages: en
• Tags: transformers, safetensors, text-generation, small-language-model, base, pretrained, hybrid, gated-deltanet

Links

• Model: https://huggingface.co/MinimaLabs/KLM-1.6
• Files: https://huggingface.co/MinimaLabs/KLM-1.6/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download MinimaLabs/KLM-1.6

Then follow the usage in its model card. Nothing in this kit runs by itself.