Zarya-0.6B

Zarya is a family of hybrid language models that combine a classic auto-regressive (AR) objective with a masked-diffusion (MDM) objective in one model. The architecture can be built on top of any autoregressive model but in this repository it uses the Qwen3 backbone.

Naming explanation: Zarya (pronounced as [zɐˈrʲa] (IPA notation), literally "Dawn" in English) is a figure from Slavic folklore — a female personification of dawn who may be considered a goddess. In various traditions, she can manifest as a single being or as two or three sisters simultaneously.

This is a research prototype.

Model Details

Model Description

Zarya is a research prototype of a family of hybrid language models that jointly learn a classic auto-regressive (AR) objective and a masked-diffusion (MDM) objective within a single model.

Two generation modes are supported, both reachable through a single model.generate(...) call: masked-diffusion (MDM) sampling and slotted-level speculative parallel decoding.

• Model type: Hybrid auto-regressive (AR) + masked-diffusion language model (DLLM); backbone Qwen3, wrapper Zarya
• Language(s) (NLP): Russian and English
• License: MIT
• Preprint: https://arxiv.org/abs/2609.19868
• Repository with training code: https://github.com/ai-forever/zarya

Zarya-0.6B details

Zarya-0.6B has the following features:

Context Length: 2048

Uses

Zarya is intended for text generation. It supports conversational fine-tuning (SFT) and classic auto-regressive pretraining.

Direct Use

Direct use is text generation (continuation of a prompt) through the model.generate(...) interface, including chat-style prompts formatted with the provided chat template. Two inference modes are available through the same generate() call. Both modes fully use the KV cache with causal attention masks.

• MDM sampling (slotted_generation=false): iterative masked-diffusion denoising with the first-hitting sampler.
• Slotted speculative decoding (slotted_generation=true): parallel slot generation with inter-slot diffusion-based selection and intra-slot autoregressive generation for a decoding speedup.

Out-of-Scope Use

The model is a research prototype. It should not be used for production decisions, safety-critical applications, or any use case where accuracy and reliability are essential without additional evaluation and safeguards. Inference performance and stability also depend on the chosen decoding hyperparameters (like slotted_generation, slot_size, serial_num_blocks, slot_threshold, token_threshold, and others).

Bias, Risks, and Limitations

This is a research prototype. The code relies on Hugging Face Transformers APIs; when upgrading versions, compatibility must be checked (tested on Transformers 5.12.1 and PyTorch 2.9.0). Inference performance and stability depend on the choice of config parameters.

How to Get Started with the Model

Use the code below to get started with the model. Loading the model and tokenizer requires trust_remote_code=True.

import torch
from transformers import AutoModel, AutoTokenizer

model_name = "ai-forever/Zarya-0.6B"

model = AutoModel.from_pretrained(model_name, trust_remote_code=True, torch_dtype=torch.bfloat16).cuda()
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)

prompt = "<|im_start|>user\nHello!<|im_end|>\n<|im_start|>assistant\n"
input_ids = tokenizer(prompt, return_tensors="pt").input_ids.cuda()

# Both modes go through model.generate(...).
# With generation_config.slotted_generation=true -> slotted speculative decoding:
out = model.generate(
    input_ids,
    max_new_tokens=256,
    do_sample=True,
    temperature=0.7,
    slot_size=16,
    serial_num_blocks=4,
    slot_threshold=0.9,
    token_threshold=0.3,
)
# Setting generation_config.slotted_generation=false -> MDM sampling instead:
# out = model.generate(input_ids, max_new_tokens=256)

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

Evaluation

LM-eval benchmarking with the lm-eval package is supported. Example run:

lm_eval run \
--tasks=gsm8k,ifeval,arc_challenge,hendrycks_math500,minerva_math500,humaneval_instruct,humaneval,mbpp,mbpp_plus,hellaswag \
--model=hf --confirm_run_unsafe_code \
--log_samples \
--apply_chat_template \
--output_path=./reports/lm-eval_results \
--model_args=pretrained=ai-forever/Zarya-0.6B,backend=causal,dtype=bfloat16,attn_implementation=sdpa,trust_remote_code=True \
--gen_kwargs slotted_generation=true,slot_size=16,serial_num_blocks=4,slot_threshold=0.9,token_threshold=0.4

Zarya-0.6B results

Measurements below were collected with varying inference parameters and on different GPUs; performance is sensitive to both, so results may differ across configurations and hardware setups.

A100, dtype=bfloat16, attn_implementation=sdpa, apply_chat_template, slotted_generation=true,slot_size=16,serial_num_blocks=4,slot_threshold=0.9,token_threshold=0.4

Hardware info: gpu_driver_cuda_version 13.2; gpu_driver_version 595.71.05;

Docker info: Torch: 2.9.0+cu128; Transformers: 5.19.0; CUDNN in torch: 91002; lm-eval == 0.4.13

A100, dtype=float16, attn_implementation=sdpa, apply_chat_template, slotted_generation=true,slot_size=16,serial_num_blocks=4,slot_threshold=0.9,token_threshold=0.4

Hardware info: gpu_driver_cuda_version 13.2; gpu_driver_version 595.71.05;

Docker info: Torch: 2.9.0+cu128; Transformers: 5.19.0; CUDNN in torch: 91002; lm-eval == 0.4.13

A100, dtype=float32, attn_implementation=sdpa, apply_chat_template, slotted_generation=true,slot_size=16,serial_num_blocks=4,slot_threshold=0.9,token_threshold=0.4

Hardware info: gpu_driver_cuda_version 13.2; gpu_driver_version 595.71.05;

Docker info: Torch: 2.9.0+cu128; Transformers: 5.19.0; CUDNN in torch: 91002; lm-eval == 0.4.13

A100, dtype=bfloat16, attn_implementation=eager, apply_chat_template, slotted_generation=true,slot_size=16,serial_num_blocks=4,slot_threshold=0.9,token_threshold=0.4

Hardware info: gpu_driver_cuda_version 13.2; gpu_driver_version 595.71.05;

Docker info: Torch: 2.9.0+cu128; Transformers: 5.19.0; CUDNN in torch: 91002; lm-eval == 0.4.13

A100, dtype=bfloat16, attn_implementation=flash_attention_2, apply_chat_template, slotted_generation=true,slot_size=16,serial_num_blocks=4,slot_threshold=0.9,token_threshold=0.4

Hardware info: gpu_driver_cuda_version 13.2; gpu_driver_version 595.71.05;

Docker info: Torch: 2.9.0+cu128; Transformers: 5.19.0; CUDNN in torch: 91002; lm-eval == 0.4.13

A100, dtype=bfloat16, attn_implementation=flash_attention_3, apply_chat_template, slotted_generation=true,slot_size=16,serial_num_blocks=4,slot_threshold=0.9,token_threshold=0.4

Hardware info: gpu_driver_cuda_version 13.2; gpu_driver_version 595.71.05;

Docker info: Torch: 2.9.0+cu128; Transformers: 5.19.0; CUDNN in torch: 91002; lm-eval == 0.4.13

A100, dtype=bfloat16, attn_implementation=flex_attention, apply_chat_template, slotted_generation=true,slot_size=16,serial_num_blocks=4,slot_threshold=0.9,token_threshold=0.4

Hardware info: gpu_driver_cuda_version 13.2; gpu_driver_version 595.71.05;

Docker info: Torch: 2.9.0+cu128; Transformers: 5.19.0; CUDNN in torch: 91002; lm-eval == 0.4.13

A100, dtype=bfloat16, attn_implementation=sdpa, apply_chat_template, slotted_generation=true,slot_size=16,serial_num_blocks=4,slot_threshold=0.9,token_threshold=0.4,max_gen_toks=2048

Hardware info: gpu_driver_cuda_version 13.2; gpu_driver_version 595.71.05;

Docker info: Torch: 2.9.0+cu128; Transformers: 5.19.0; CUDNN in torch: 91002; lm-eval == 0.4.13

H100, dtype=bfloat16, attn_implementation=sdpa, apply_chat_template, slotted_generation=true,slot_size=16,serial_num_blocks=4,slot_threshold=0.9,token_threshold=0.4

Hardware info: gpu_driver_cuda_version 13.0; gpu_driver_version 580.105.08;

Docker info: Torch: 2.9.0+cu128; Transformers: 5.19.0; CUDNN in torch: 91002; lm-eval == 0.4.13

---

Citation

If you find our work helpful, please consider citing (citation will be updated after peer-reviewed publication):

@misc{sinev-etal-2026-Zarya,
  author        = {Sinev, Leonid and Koziev, Ilya and Leshchuk, Vladislav},
  title         = {Zarya: A Hybrid Autoregressive--Masked Diffusion Language Model with Flexible Training and Dual-Mode Inference},
  year          = {2026},
  archiveprefix = {arXiv},
  eprint        = {2609.19868},
  primaryclass  = {cs.CL},
  url           = {https://arxiv.org/abs/2609.19868},
}