GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
DhruvParmar051
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
DhruvParmar051
Model Tree
Input Modalities
Output Modalities
Supported Functionality
| Architecture | decoder-only transformer, LlamaForCausalLM (RMSNorm, RoPE, SwiGLU, tied embeddings) |
| Parameters | 11,798,784 (3,410,176 non-embedding) |
| Layers / d_model / heads / d_ff | 4 / 256 / 4 / 768 |
| Context length | 1024 tokens |
| Vocabulary | 32,768 (byte-level BPE) |
| RoPE theta / norm eps | 10000 / 1e-05 |
| Weights | fp32 safetensors |
| Data | TinyStories (CDLA-Sharing-1.0) |
| Tokens seen | 294,912,000 (9,000 steps x 32,768 tokens) |
| Optimizer | AdamW, peak LR 0.001, min LR 0.0001, 200 warmup steps, cosine decay, weight decay 0.1 |
| Precision | bf16-autocast |
| Hardware | Apple M5 Pro laptop (20-core GPU, 48 GB unified memory), PyTorch MPS backend |
| Wall time | 2.48 h |
| Final train / val loss | 1.5620 / 1.5725 (val: training-time estimate) |
Source: EXP-012.
Full validation split, non-overlapping 1024-token windows (2,073,600 target tokens, 8,110,848 bytes), bf16-autocast:
| Loss (nats/token) | Perplexity | Bits per byte |
|---|---|---|
| 1.5641 | 4.78 | 0.5769 |
Bits per byte is the only one of these comparable across tokenizers. Source: EXP-014.
lm-evaluation-harness 0.4.13, 0-shot, float32, <|endoftext|> prepended, on the pinned benchmark files. ± is the binomial standard error. "Clean" excludes examples that share a 13-gram with the training data and examples too short to check.
| Benchmark | Metric | All | Clean | Contaminated | Chance |
|---|---|---|---|---|---|
| HellaSwag | acc_norm | 25.9% ± 0.4 (n=10,042) | 25.9% ± 0.4 (n=10,002) | 22.5% ± 6.6 (n=40) | 25.0% |
| ARC-Easy | acc | 27.0% ± 0.9 (n=2,376) | 27.2% ± 0.9 (n=2,278) | 4.5% ± 4.4 (n=22) | 25.0% |
| PIQA | acc_norm | 53.0% ± 1.2 (n=1,838) | 53.5% ± 1.2 (n=1,682) | 33.3% ± 27.2 (n=3) | 50.0% |
| LAMBADA (OpenAI) | acc | 7.8% ± 0.4 (n=5,153) | 7.8% ± 0.4 (n=5,153) | n/a | n/a |
Sources: EXP-017 (scores), EXP-016 (contamination).
No comparisons with other models are made: published scores use different training data, token counts, tokenizers and prompts, so they would not be like for like.
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer tok = AutoTokenizer.from_pretrained("DhruvParmar051/syrinx-10m")model = AutoModelForCausalLM.from_pretrained("DhruvParmar051/syrinx-10m", dtype=torch.float32)# Training documents always followed <|endoftext|>, so start prompts with it.ids = tok(tok.bos_token + "Once upon a time", return_tensors="pt")out = model.generate(**ids) # uses the shipped generation_configprint(tok.decode(out[0], skip_special_tokens=True))Generating short, simple children's stories, and studying a small language model trained end to end on one laptop. It is a base model: it continues text and does not follow instructions.
| Item | Experiment | File | SHA-256 |
|---|---|---|---|
| Training | EXP-012 | experiments/10m-a-train-2026-10-12.txt | 830997730463c63d95385a5db7a6d4cecb7c6044ed61c863aaed8524282a4808 |
| Evaluation | EXP-014 | experiments/eval-10m-a-step9000-bf16-autocast-2026-10-13.json | 36de4dffede0a71036517d4c34f964b6fd9931d247fb597aff6559b969ba03e4 |
| Export parity | EXP-015 | experiments/export-10m-a-step9000-parity-2026-10-13.json | e78bcefc5b370ab4b22090031901274f41c88ae027844e38e2721249a76fdc9f |
| Contamination | EXP-016 | experiments/contamination-2026-10-14.json | 839e2d86e90a4d68cb1bb6c1213ea6e61252c1ea7c6d76cb8df644c2fc64bda6 |
| Benchmarks | EXP-017 | experiments/benchmarks-10m-a-2026-10-14.json | e3eb81340e6b3c35cf4ef58f6c5dd0dc235e385118d913dccbe086f78faa3ef4 |
| Weights | model.safetensors | 5e50d4d295bc69a052ba3d024413946fa5e801bdefc725a8634f4fe44bd50f8a |
Export parity: max absolute logit difference 0 against the Syrinx implementation (tolerance 0.0001).
Code and weights: Apache-2.0.
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer tok = AutoTokenizer.from_pretrained("DhruvParmar051/syrinx-10m")model = AutoModelForCausalLM.from_pretrained("DhruvParmar051/syrinx-10m", dtype=torch.float32)# Training documents always followed <|endoftext|>, so start prompts with it.ids = tok(tok.bos_token + "Once upon a time", return_tensors="pt")out = model.generate(**ids) # uses the shipped generation_configprint(tok.decode(out[0], skip_special_tokens=True))