Is Udgam-384M state of the art?Pending benchmarks.Announced Oct 10, 2026; no independent results yet.

The details so far

Lab

Udgam Labs

Announced

Oct 10, 2026

API

Not yet

Open weights

Published

Parameters

384M (384.4M; 20 layers, hidden size 1,024)

Architecture

Decoder-only transformer compatible with Gemma 4 (Gemma4ForCausalLM), sliding-window plus global attention

Training

Trained from scratch (random initialization) on roughly 8B tokens of licensed public and synthetic text, then supervised fine-tuning for chat and tool use

Context

4,096 tokens (2,048 for the base model)

Language

English

Formats

safetensors, GGUF (Q8_0 410 MB; Q4_K_M 238 MB for phones) and MLX (4-bit build 216 MB); runs offline in Ollama, LM Studio, llama.cpp, MLX and transformers

Release

Udgam-384M-Chat plus Udgam-384M-Base and Udgam-384M-Base-4K on Hugging Face; training code and data not released

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.

• 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: on Udgam's own 274-item held-out sets (which follow this model's prompt conventions), 75% right call/no-call decisions on relevance, 56% schema-valid JSON for unseen schemas and 85.9% right shell behaviour.

Lab claim: BFCL v3 simple 60, IFEval prompt-level strict 26.1; Udgam reports Granite 4.0 350M (90.7 / 47) and Qwen3-0.6B (80 / 58) ahead on both in the same harness.

Lab claim: XSTest declines 54% of clearly harmful prompts and over-refuses 15% of safe ones; released as not safety-hardened.

Lab claim: 444 tokens/s generation at Q4_K_M on an Apple M6 Mac mini GPU.

Sources