GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
altslate
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
altslate
Model Tree
Input Modalities
Output Modalities
Supported Functionality
Evaluated with EleutherAI lm-evaluation-harness:
| Metric | Score |
|---|---|
| BLiMP (acc) | 78.14% |
| ARC-Easy (acc) | 51.43% |
| WikiText-2 (byte perplexity ↓) | 2.04 |
| WikiText-2 (bits/byte ↓) | 1.03 |
| Validation perplexity ↓ | 21.6 |
Reproduce:
bash
lm_eval --model hf \--model_args pretrained=altslate/JugnuLM-53M,dtype=bfloat16 \--tasks blimp,arc_easy,wikitext --batch_size auto
| Params | 53.5M |
| Base architecture | Qwen3 (Llama-style + built-in QK-Norm) |
| Layers | 8 |
| Hidden size | 512 |
| Heads / KV heads | 8 / 4 (grouped-query attention) |
| Context | 2048 |
| Position | RoPE |
| Norm / FFN | RMSNorm / SwiGLU |
| Embeddings | tied |
| Tokenizer | SmolLM2 (49,152 vocab) |
HuggingFaceFW/fineweb-edu (sample-10BT), ~12B tokens seen, packed into 2048-token blockspython
from transformers import AutoModelForCausalLM, AutoTokenizertok = AutoTokenizer.from_pretrained("altslate/JugnuLM-53M")model = AutoModelForCausalLM.from_pretrained("altslate/JugnuLM-53M")ids = tok("The capital of France is", return_tensors="pt").input_idsout = model.generate(ids, max_new_tokens=30, repetition_penalty=1.3)print(tok.decode(out[0], skip_special_tokens=True))
Research / leaderboard entry. This is a base model — not instruction-tuned or aligned. Expect fluent English continuation and reasonable factual recall for its size, but limited multi-step reasoning, occasional repetition, and hallucination. Not for production use.
Trained by AltSlate Labs. Recipe and training code: https://github.com/AltSlate-Labs/jugnu