Sutra-1.3B-Chat

Sutra-1.3B

<!-- The counter above this card is fixed to a rolling 30 days and cannot be switched. This one reads downloadsAllTime from the Hub API on every page load, so the two together show both the recent rate and the total since launch. -->

A 1.32B-parameter Mixture-of-Experts language model pretrained from scratch in pure PyTorch, with Multi-head Latent Attention (MLA) and DeepSeek-style auxiliary-loss-free expert routing — implemented from the papers, not adapted from an existing codebase.

Own BPE tokenizer, own streaming data pipeline, own training loop. No pretrained weights, no transformers Trainer, no reference implementation anywhere in the stack.

If you are here to learn how a MoE model is trained end to end, the code is the point: github.com/Abhisingh18/Sutra-1.3B-Model carries the tokenizer, data prep, pretraining, SFT, DPO, evaluation and serving — plus the six silent bugs that cost the most time, written up in full.

Try it without installing anything

sutra-1-3-b-model-15co.vercel.app — a live chat demo running these weights, with web search and document upload on top.

The inference widget on this page stays empty on purpose: Inference Providers serve standard transformers architectures, and this is a custom MoE with Multi-head Latent Attention. Same reason AutoModelForCausalLM does not work here — the architecture code ships with the weights, not with transformers.

Quick start

from transformers import AutoModelForCausalLM
from tokenizers import Tokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "Abhisingh-18/Sutra-1.3B-Chat", trust_remote_code=True,
    dtype=torch.bfloat16).eval()
tok = Tokenizer.from_file("tokenizer.json")

prompt = ("<|begin_of_text|><|user|>\nWhat is the capital of France?"
          "<|end_turn|>\n<|assistant|>\n")
ids = torch.tensor([tok.encode(prompt).ids])
out = model.generate(ids, max_new_tokens=64, min_new_tokens=12, do_sample=False)
print(tok.decode(out[0].tolist()[ids.shape[1]:]))

trust_remote_code=True is required: the architecture (MoE + Multi-head Latent Attention) ships with the weights rather than with transformers. The remote code is verified to produce logits identical to the reference implementation to 7e-06 in fp32.

Or without transformers at all:

pip install torch tokenizers huggingface_hub
wget https://huggingface.co/Abhisingh-18/Sutra-1.3B-Chat/resolve/main/inference.py
python inference.py "Explain photosynthesis in three sentences."

That downloads the weights (5.3 GB) and the architecture code and runs. It works on CPU — because only 0.28B parameters are active per token, CPU generation runs at about 10 tokens/second on 2 cores.

Interactive:

python inference.py

In Python:

from inference import build, generate
model, mcfg, tok, device = build()
print(generate(model, mcfg, tok, device, "What is machine learning?"))

Note: min_new_tokens matters. Left to itself this model often emits <|end_turn|> immediately and returns an empty string.

Repository layout

├── model.safetensors        default weights — the DPO stage, 5.3 GB
├── config.json              architecture
├── tokenizer.json           48k BPE, English + Devanagari
├── inference.py             run the model
├── src/                     model code (custom arch — see note above)
└── checkpoints/             every training stage, for comparison
    ├── base_pretrained.pt
    ├── sft_epoch_0.pt
    ├── sft_epoch_1.pt
    ├── sft_epoch_2.pt
    ├── dpo_epoch_0.pt
    └── dpo_epoch_0.int8.pt   quantized — 1.42 GB

The root holds what you need to run the model. checkpoints/ is the archive of each stage, so you can hear the difference each one made instead of taking it on faith.

Optimizer state is stripped from all of them, so each is 5.3 GB rather than 15.8 GB. They are for inference, not for resuming training.

To load a different stage:

SUTRA_CKPT=checkpoints/base_pretrained.pt python inference.py "The capital of France is"

Architecture

Training

Pretraining held-out perplexity 15.00; SFT held-out perplexity 5.49.

Every dataset behind all three stages — with its exact config, split, text field and token share, plus the prep script — is published as Sutra-1.3B-Data, so the corpus can be rebuilt from the original sources.

Evaluation

Log-likelihood scoring, 500 examples per task, length-normalised accuracy.

Two things worth reading honestly here. ARC-easy and PIQA sit well above chance, so the model learned real commonsense and not just fluent grammar. WinoGrande sits at chance, which is the clearest signal of what 0.28B active parameters cannot buy: the pronoun-resolution reasoning that task measures never appeared.

DPO's held-out preference accuracy came out at 47.5% against a 50% baseline, so the alignment stage did not generalise — the 66% reported during training was measured on training batches. The SFT and DPO checkpoints perform about equally.

INT8 quantized weights

checkpoints/dpo_epoch_0.int8.pt is the same model at 1.42 GB instead of 2.64 GB (1.62 GB resident on GPU). Measured on all four tasks, 500 examples each:

At 500 examples, 0.2 points is a single example — the difference is noise.

Weight-only, symmetric, per-output-channel: each row of each weight matrix carries its own scale, because rows differ in range by orders of magnitude. Activations stay bf16, and lm_head is left unquantized. This saves memory, not time — the weight is dequantized before the matmul, so arithmetic runs at the same speed.

SUTRA_CKPT=checkpoints/dpo_epoch_0.int8.pt python inference.py "Explain gravity."

Not to be confused with DeepSeek's FP8 work, which is a training technique. This is post-training quantization, applied to finished weights.

Limitations

Trained on 18B tokens — roughly 500x less than comparable 1B models such as Llama 3.2 1B (9T tokens). Concretely:

• Writes fluent English and follows formatting instructions well
• Does not reliably recall facts, and states wrong ones confidently
• Does not do multi-step reasoning or write working code
• Sensitive to phrasing — a typo or a terse prompt derails it, where a larger model would recover

Pair it with retrieval for anything knowledge-dependent.

License

Apache 2.0.