Skip to content

Speck1-140M / Base model

Speck1-140M

An open English base language model for local inference, fine-tuning, and architecture research.

140.7M

Parameters

5B

Pretraining tokens

2,048

Validated context

MIT

Open license

01 / Overview

About Speck1

Base model

Use it for text completion, fine-tuning, or language-model research. For chat-style prompts, use an Instruct checkpoint.

Hybrid network

Grouped-query attention is combined with gated causal convolution to reduce runtime state while retaining global attention.

Local formats

Available as BF16 Safetensors for Transformers and quantized GGUF files for llama.cpp-compatible applications.

02 / Comparison

Size, speed, and memory

Speck1 is shown alongside four similarly sized models. Switch views to compare evaluation score, training budget, inference speed, and memory.

Open SLM Intelligence Index

Higher is better

Score / 30

SmolLM2-135M27.13
GPT-X2.5-135M25.17
BananaMind-2-Pro24.96
Supra2-100M-Base19.41
Speck1-140M18.15

Pretraining tokens

Lower means fewer tokens

Log scale

SmolLM2-135M~2T
BananaMind-2-Pro100B
GPT-X2.5-135M75B
Supra2-100M-Base30B
Speck1-140M5B

Scores use a pinned Open SLM configuration. Speck is not yet a published leaderboard row, and training budgets are not compute-matched.

03 / Before use

Practical notes

Context length

Training and validation used sequences up to 2,048 tokens. The 4,096-token configuration has not been validated.

Capabilities

At 140M parameters, knowledge, reasoning, arithmetic, coding, and multilingual performance are limited.

Safety

The model has no dedicated safety alignment and can produce incorrect, biased, or unsafe text. It is not intended for high-stakes use.

Methodology

Sources

Model details and local performance measurements come from the public model card and repository. Quality scores use a pinned Open SLM evaluation configuration; Speck is not currently a published row in the linked leaderboard snapshot.