Is Laghu state of the art?Pending benchmarks.Announced Oct 22, 2026; no independent results yet.
Udgam Labs' 152M-parameter English chat model with a thinking on/off switch, trained from scratch with a $146 main pretraining run and released fully open: weights, late checkpoints, fine-tuning data, evaluation outputs and code. The 163 MB model runs in the browser. Apache-2.0.
The details so far
- Lab
- Udgam Labs
- Announced
- Oct 22, 2026
- API
- Not yet
- Open weights
- Published
- Parameters
- 152M (151.8M; 28 layers, width 640)
- Architecture
- Gemma 4 text architecture: 10 query and 2 key-value heads; three 512-token sliding-window layers then one global layer, repeated
- Training
- Trained from scratch with Udgam's own 32,768-token tokenizer: about 70B tokens read in pretraining at 2K, a 1B-token stretch at 32K, then fine-tuning on 41,467 conversations
- Cost
- Main pretraining run $146.26 (8× A100 80GB, 13.2 hours); about $194 for the whole project
- Context
- 16K tokens for use (stretched at 32K; recall fails beyond about 16K)
- Modes
- Thinking on or off; Udgam recommends off except for short arithmetic
- Release
- Laghu, Laghu-Base, Laghu-Base-32K, Laghu-Checkpoints (late training states only; earlier checkpoints were lost), GGUF and MLX builds, the laghu-sft and laghu-evals datasets, and code on GitHub (udgamlabs/laghu)
- License
- Apache-2.0 for weights and code; nanochat-derived files keep their MIT notice
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: level with SmolLM2-135M-Instruct on judged chat quality (51.0% win rate) and multi-turn follow-ups (41.7% vs 37.5%); behind Falcon-H1-Tiny-90M-Instruct on all three of Udgam's pre-registered chat tests, so it missed Udgam's own release bar.
Lab claim: GSM8K 7.9% with thinking on, IFEval prompt-level strict 26.1.
Lab claim: Laghu-Base scores 0.919 bits per byte on text published in September–October 2026, against 0.941 for SmolLM2-135M and 0.979 for Falcon-H1-Tiny-90M-Base.
Sources
LABLaghu release post (Udgam Labs blog)udgamlabs.com/blog/laghu · Oct 22, 2026LABLaghu model page (Udgam Labs)udgamlabs.com/models/laghu · Oct 22, 2026REPOudgamlabs/Laghu model card (Hugging Face)huggingface.co/udgamlabs/LaghuREPOudgamlabs/laghu (GitHub)github.com/udgamlabs/laghuLABUdgam Labs on Xx.com/udgamlabs/status/2113272462062637075 · Oct 22, 2026Tracking continues until an independent board lists it.