Qwen3¶

Weights: pretrained Keras weights live on Hugging Face under zeromodels/<variant> (each repo carries kf_config.json and tokenizer.json plus the Keras weights: model.weights.h5, or a sharded model.weights.json + shards for the larger checkpoints). Load with from_weights("zeromodels/<variant>").

Alibaba's Qwen3 dense decoder-only LLM, ported to pure Keras 3. It drops Qwen2's QKV biases and adds QK-RMSNorm (queries and keys are normalized before attention), keeping RMSNorm, SwiGLU and grouped-query attention otherwise.

Links:

HF collection: Qwen3
Paper: Qwen3 Technical Report (arXiv:2505.09388)
HF docs: transformers/model_doc/qwen3

Variants¶

Preconverted, bf16 weights are hosted under zeromodels/. Load with from_weights("zeromodels/<variant>"); the -base suffix marks the base (non-instruction-tuned) checkpoints, and the -2507 variants are the July 2025 refresh (same 4B architecture, longer context). Qwen3 is Apache 2.0. The MoE sizes live on qwen3_moe.md.

Variant	Hub
`qwen3-0.6b`	`zeromodels/qwen3-0.6b`
`qwen3-0.6b-base`	`zeromodels/qwen3-0.6b-base`
`qwen3-1.7b`	`zeromodels/qwen3-1.7b`
`qwen3-1.7b-base`	`zeromodels/qwen3-1.7b-base`
`qwen3-4b`	`zeromodels/qwen3-4b`
`qwen3-4b-base`	`zeromodels/qwen3-4b-base`
`qwen3-4b-instruct-2507`	`zeromodels/qwen3-4b-instruct-2507`
`qwen3-4b-thinking-2507`	`zeromodels/qwen3-4b-thinking-2507`
`qwen3-8b`	`zeromodels/qwen3-8b`
`qwen3-8b-base`	`zeromodels/qwen3-8b-base`
`qwen3-14b`	`zeromodels/qwen3-14b`
`qwen3-14b-base`	`zeromodels/qwen3-14b-base`
`qwen3-32b`	`zeromodels/qwen3-32b`

Upstream Qwen safetensors also load directly via the hf: prefix, e.g. from_weights("hf:Qwen/Qwen3-4B-Instruct-2507"), which converts them in process (pass cache_converted=True to keep the result). See Loading Weights.

API¶

`Qwen3Model`¶

The decoder backbone, no LM head. Returns {"last_hidden_state": (batch, seq, embed_dim)}.

Arg	Default	Meaning
`vocab_size`	`151936`	token vocabulary size
`embed_dim`	`1024`	model width
`mlp_dim`	`3072`	MLP inner width
`num_layers`	`28`	decoder blocks
`num_heads`	`16`	query heads
`num_kv_heads`	`8`	key/value heads (GQA)
`head_dim`	`128`	per-head width
`norm_eps`	`1e-06`	RMSNorm epsilon
`rope_theta`	`1000000.0`	rotary base frequency
`tie_embeddings`	`True`	reuse the embedding matrix as the LM head

`Qwen3TextGenerate`¶

Qwen3Model plus a (tied) LM head. Returns {"logits": (batch, seq, vocab_size)} and adds .generate(). Same constructor arguments as Qwen3Model.

generate(
    input_ids,
    attention_mask=None,
    max_new_tokens=None,
    eos_token_id=None,
    sampler=None,
    seed=None,
    **prefill_inputs,
)

Arg	Default	Meaning
`input_ids`	required	`(batch, seq)` token ids
`attention_mask`	`None`	`(batch, seq)` 1 = keep, 0 = padding
`max_new_tokens`	`None`	tokens to generate
`eos_token_id`	`None`	stop token (defaults to the tokenizer's)
`sampler`	`None`	sampling strategy; greedy when unset
`seed`	`None`	seed for stochastic samplers

`Qwen3Tokenizer`¶

Tokenizer on the tokenizers backend.

Qwen3Tokenizer(hf_id=None, tokenizer_file=None)

Arg	Default	Meaning
`hf_id`	`None`	Hub repo to pull the tokenizer files from
`tokenizer_file`	`None`	explicit path to a `tokenizer.json`

Calling it returns {"input_ids", "attention_mask"}, padded across the batch. It accepts a plain string, a list of strings (a batch), or a chat-message list, which is routed through apply_chat_template automatically. Decode with .decode(ids) for one sequence or .batch_decode(ids) for a batch.

End-to-end example¶

Single input¶

import os

os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from zeromodels.models.qwen3 import Qwen3TextGenerate, Qwen3Tokenizer

model = Qwen3TextGenerate.from_weights("zeromodels/qwen3-0.6b")
tokenizer = Qwen3Tokenizer.from_weights("zeromodels/qwen3-0.6b")

inputs = tokenizer(
    [{"role": "user", "content": "Explain rotary embeddings in one sentence."}]
)
outputs = model.generate(**inputs, max_new_tokens=64)

print(tokenizer.decode(outputs[0]))

Batch¶

Pass a list of strings. The tokenizer pads them and generate runs the batch together:

prompts = [
    "The capital of France is",
    "In one sentence, what is a transformer?",
    "Write a haiku about GPUs.",
]
inputs = tokenizer(prompts)  # {"input_ids": (3, seq), "attention_mask": (3, seq)}
outputs = model.generate(**inputs, max_new_tokens=64)

for text in tokenizer.batch_decode(outputs):
    print(text)

Backbone only¶

from zeromodels.models.qwen3 import Qwen3Model

backbone = Qwen3Model.from_weights("zeromodels/qwen3-0.6b")
hidden = backbone(inputs)["last_hidden_state"]  # (batch, seq, embed_dim)

Loading from the Hub¶

Any Hub repo with this architecture works via the hf: prefix, including community fine-tunes:

model = Qwen3TextGenerate.from_weights("hf:Qwen/Qwen3-0.6B")

Lower memory¶

Larger checkpoints load in bf16 or weight-only quantized. See quantization.md:

model = Qwen3TextGenerate.from_weights(
    "zeromodels/qwen3-0.6b", quantization="int8", load_dtype="bfloat16"
)