# Is Gargi SOTA?

Pending benchmarks. Announced Oct 17, 2026; no independent results yet.

As of Nov 3, 2026, Gargi is pending benchmarks on isitsota: announced Oct 17, 2026, with no independent results on the boards we track yet.

Udgam Labs' 384M-parameter English language model, trained from scratch on about 17 billion tokens, with optional step-by-step thinking. Trained at 32K context with reliable recall up to about 16K tokens. Apache-2.0, with GGUF and MLX builds for Ollama, llama.cpp and Apple silicon.

Canonical: https://isitsota.com/m/gargi
Last updated: 2026-10-24
Tracking since: 2026-10-24

## Announced facts

| Fact | Value |
| --- | --- |
| Lab | Udgam Labs |
| Announced | Oct 17, 2026 |
| API | Not yet |
| Open weights | Published |
| Parameters | 384M |
| Architecture | Gemma 4-style decoder, sliding-window plus global attention |
| Training | Trained from scratch on about 17B tokens of openly licensed or public-domain text plus synthetic text from open models |
| Context | Trained at 32K tokens; Udgam advises inputs up to about 16K |
| Modes | Thinking on (step-by-step reasoning before the answer) or off |
| Release | Gargi (chat) and Gargi-Base on Hugging Face, plus GGUF (Q8_0 410 MB, Q4_K_M 238 MB) and MLX builds; runs in Ollama and llama.cpp |
| Language | English |
| License | Apache-2.0 (weights and tokenizer) |

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.

- [x] Publicly announced with a source
- [x] 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: Gargi-Base scores 62.8 ARC-Easy, 32.3 ARC-Challenge and 54.6 WinoGrande against 64.7, 32.4 and 56.1 for Qwen2.5-0.5B in the same harness; clearly behind on MMLU (25.5), HellaSwag and PIQA.
- Lab claim: GSM8K 36.0 with thinking on, 32.9 with thinking off (0-shot, chat template).
- Lab claim: needle-in-a-haystack recall 100% at 8K, 93% at 16K, 90% at 20K and 0% at 32K tokens.
- Lab claim: not safety-tuned beyond a small set of refusal examples.

## Sources

- LAB: [Gargi release post (Udgam Labs blog)](https://udgamlabs.com/blog/gargi) (2026-10-17)
- LAB: [Gargi model page (Udgam Labs)](https://udgamlabs.com/models/gargi) (2026-10-17)
- REPO: [udgamlabs/Gargi model card (Hugging Face)](https://huggingface.co/udgamlabs/Gargi)
- REPO: [udgamlabs/gargi (GitHub)](https://github.com/udgamlabs/gargi)

Tracking continues until an independent board lists it.

## Related

- Lab: [Udgam Labs](https://isitsota.com/labs/udgam-labs.md)
- All pending releases: https://isitsota.com/releases.md
- JSON: https://api.isitsota.com/v1/models/gargi