UDGAM LABS · Language model · 278M parameters, Telugu, base · GGUF

Vemana-Base-GGUF

Vemana-Base, a 278M-parameter Telugu language model trained from scratch, as GGUF files for llama.cpp and for running in a web browser. It is a base model: it continues Telugu text and does not follow instructions. Results, data and limits are on the model card.

File	Size	Use
Vemana-Base-Q8_0.gguf	297 MB	default; in the browser; matches the original model closely
Vemana-Base-F16.gguf	558 MB	the original model's numbers in 16-bit floats

Quick start

llama.cpp:

llama-completion -hf udgamlabs/Vemana-Base-GGUF:Q8_0 -p "వర్షాకాలంలో" -n 80 --temp 0.8 --top-p 0.95 -c 1024

In a web browser: try it at udgamlabs.com/models/vemana. To build your own page (llama.cpp compiled to WebAssembly with wllama; nothing leaves the device): load Vemana-Base-Q8_0.gguf with n_ctx: 1024 and send plain Telugu text as the prompt; wllama adds the <s> token itself. On a Mac mini (Chromium) it writes about 100 tokens per second.

Keep prompt and output within 1,024 tokens, the length it was trained at.

How these files were made and checked

• Converted from Vemana-Base with llama.cpp's convert_hf_to_gguf.py (Q8_0 and F16).
• Two metadata changes, weights untouched. llama.cpp's Gemma 4 code needs at least one sliding-window layer, and Vemana uses full attention everywhere, so layer 0 is marked sliding with a window of 4,096 tokens (the file's whole context) and the same rotary and head settings as the other layers: the computation is unchanged. The tokenizer is stored as Vemana's SentencePiece model, which llama.cpp then splits exactly like the original.
• Checked against the original model (transformers, float32) on 27 October 2026 [measured by us]:
  • tokenization identical on 40 of 40 Telugu Wikipedia pages;
  • perplexity on the same 4,080 tokens: original 15.764, F16 15.764, Q8_0 15.765;
  • greedy continuations of four prompts, 40 tokens: F16 identical on all four; Q8_0 identical on three and differs after a few words on one (8-bit rounding);
  • in the browser (wllama 3.6.1), prompt token counts and two greedy continuations identical.
• Ollama and LM Studio were not tested.

The scripts are in github.com/udgamlabs/vemana.

Licence

Apache-2.0. Copyright 2026 Udgam Labs LLP. Training-data credits in NOTICE.

---

About Udgam Labs. Udgam Labs is an independent AI research lab. We work from first principles: before we build, we understand why things work, whether we are training a model, fine-tuning one or applying AI to a new problem. Udgam (उद्गम) is Sanskrit for origin. udgamlabs.com