Qwen3.5 (Qwen3-Next)¶

Alibaba's Qwen3.5 hybrid-attention LLM, ported to pure Keras 3. It interleaves Gated-DeltaNet linear-attention layers with periodic full-attention layers (full_attention_interval), keeping partial rotary embeddings and QK-RMSNorm on the full-attention path.

Variants¶

Load any of these with from_weights("<variant>").

Variant	Hub
`qwen3.5-0.8b`	`Qwen/Qwen3.5-0.8B`
`qwen3.5-0.8b-base`	`Qwen/Qwen3.5-0.8B-Base`
`qwen3.5-2b`	`Qwen/Qwen3.5-2B`
`qwen3.5-2b-base`	`Qwen/Qwen3.5-2B-Base`
`qwen3.5-4b`	`Qwen/Qwen3.5-4B`
`qwen3.5-4b-base`	`Qwen/Qwen3.5-4B-Base`
`qwen3.5-9b`	`Qwen/Qwen3.5-9B`
`qwen3.5-9b-base`	`Qwen/Qwen3.5-9B-Base`
`qwen3.5-27b`	`Qwen/Qwen3.5-27B`

API¶

`Qwen3_5Model`¶

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

Arg	Default	Meaning
`vocab_size`	`248320`	token vocabulary size
`embed_dim`	`1024`	model width
`mlp_dim`	`3584`	MLP inner width
`num_layers`	`24`	decoder blocks
`num_heads`	`8`	query heads
`num_kv_heads`	`2`	key/value heads (GQA)
`head_dim`	`256`	per-head width
`norm_eps`	`1e-06`	RMSNorm epsilon
`rope_theta`	`10000000.0`	rotary base frequency
`partial_rotary_factor`	`0.25`	fraction of each head that gets rotated
`tie_embeddings`	`True`	reuse the embedding matrix as the LM head
`full_attention_interval`	`4`
`linear_conv_kernel_dim`	`4`
`linear_key_head_dim`	`128`
`linear_value_head_dim`	`128`
`linear_num_key_heads`	`16`
`linear_num_value_heads`	`16`

`Qwen3_5Generate`¶

Qwen3_5Model plus a (tied) LM head. Returns {"logits": (batch, seq, vocab_size)} and adds .generate(). Same constructor arguments as Qwen3_5Model.

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

`Qwen3_5Tokenizer`¶

Tokenizer on the tokenizers backend.

Qwen3_5Tokenizer(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 kerasformers.models.qwen3_5 import Qwen3_5Generate, Qwen3_5Tokenizer

model = Qwen3_5Generate.from_weights("qwen3.5-0.8b")
tokenizer = Qwen3_5Tokenizer.from_weights("qwen3.5-0.8b")

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 kerasformers.models.qwen3_5 import Qwen3_5Model

backbone = Qwen3_5Model.from_weights("qwen3.5-0.8b")
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 = Qwen3_5Generate.from_weights("hf:Qwen/Qwen3.5-0.8B")

Lower memory¶

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

model = Qwen3_5Generate.from_weights(
    "qwen3.5-0.8b", quantization="int8", low_memory=True, load_dtype="bfloat16"
)