Is Vemana state of the art?Pending benchmarks.Announced Oct 27, 2026; no independent results yet.
Udgam Labs' 278M-parameter Telugu base model, trained from scratch with its own Telugu tokenizer and released under Apache-2.0. It continues Telugu text (it does not follow instructions) and runs on a laptop CPU or in a browser tab.
The details so far
- Lab
- Udgam Labs
- Announced
- Oct 27, 2026
- API
- Not yet
- Open weights
- Published
- Parameters
- 278M (278,214,144)
- Type
- Base model: continues Telugu text; no instruction or safety tuning
- Architecture
- Gemma 4 text layout (Gemma4ForCausalLM): 18 layers, width 1,024, 8 query and 2 key-value heads, full attention, tied embeddings
- Tokenizer
- Its own 32,000-token Telugu tokenizer with byte fallback
- Context
- 1,024 tokens (the length it was trained at)
- Training data
- 5.1B unique tokens of public Telugu web text from Sangraha (AI4Bharat), FineWeb-2 and HPLT 3.0, read about 2.35 times (12.0B token positions)
- Training run
- From random initialization on one H100 in about 16 hours, Oct 26-27, 2026, streamed live
- Release
- Vemana-Base weights, every 10% checkpoint (Vemana-Checkpoints), GGUF Q8_0 (297 MB) and F16 (558 MB), and the evaluation code on GitHub
- Browser demo
- Runs fully in the browser tab on udgamlabs.com (297 MB download, text stays on the device); also a Hugging Face Space
- License
- Apache-2.0
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: about $69 of GPU time for the training run.
Lab claim: 0.361 bits per byte on 416 Telugu Wikipedia pages from Sep-Oct 2026, ahead of every other open base model under 1B it tested, including nanochat-telugu-560M (0.383); Sarvam-1 (2.5B) scores 0.356.
Lab claim: 0-shot Belebele-te 25.1%, MILU-te 25.3%, IndicXNLI-te 35.0%, IndicCOPA-te 58.4%, which the lab reports as at or near chance.
Lab claim: about 65 tokens per second on a Mac mini CPU.
Sources
LABVemana, a small Telugu model trained live in public (Udgam Labs blog)udgamlabs.com/blog/vemana · Oct 27, 2026LABVemana model page with in-browser demo (Udgam Labs)udgamlabs.com/models/vemana · Oct 27, 2026PAPERVemana: a 278M Telugu language model trained from scratch (technical report)udgamlabs.com/research/vemana · Oct 27, 2026REPOudgamlabs/Vemana-Base on Hugging Facehuggingface.co/udgamlabs/Vemana-Base · Oct 27, 2026REPOudgamlabs/vemana on GitHubgithub.com/udgamlabs/vemana · Oct 27, 2026LABJitesh K Prasad (Udgam Labs) on X: Vemana releasex.com/jiteshkprasad/status/2115025372723151272 · Oct 27, 2026Tracking continues until an independent board lists it.