SAVRN
Search Contact SAVRN

Open-weight model · Text generation

dQwen3.5-2B-Base

by IFML UT-IFML/dQwen3.5-2B-Base

A masked diffusion language model adapted from Qwen3.5-2B. The backbone is hybrid: only its attention layers are made bidirectional, and the Gated DeltaNet layers stay causal. This is a base model, with no instruction tuning.

Parameters1.9B
Context262,144
Weights3.8 GB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads463

Runs On

What it takes to serve dQwen3.5-2B-Base (1.9B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.

PrecisionWeightsMemory neededCheapest setupPer hourAlso fits
16-bit 3.8 GB 4.5 GB 1x MI300X (192 GB)
Vultr		 $1.85 1x H100 $1.99 · 1x MI325X $2.00
8-bit 1.9 GB 2.3 GB 1x MI300X (192 GB)
Vultr		 $1.85 1x H100 $1.99 · 1x MI325X $2.00
4-bit 0.9 GB 1.1 GB 1x MI300X (192 GB)
Vultr		 $1.85 1x H100 $1.99 · 1x MI325X $2.00

Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Sep 18, 2026.

Model Card

By IFML, published under apache-2.0, revision 297e02d4a8f0.

A masked diffusion language model adapted from Qwen3.5-2B. The backbone is hybrid: only its attention layers are made bidirectional, and the Gated DeltaNet layers stay causal. This is a base model, with no instruction tuning. Paper: dQwen3.5: Hybrid-Attention Diffusion Language Models. Code: https://github.com/AntonXue/dQwen Needs a CUDA GPU and transformers>=5.13 (tested with torch 2.7.1+cu128, flash-linear-attention 0.5.1). generate decodes the whole canvas at once, committing positions above a confidence threshold (tau=0.9); pass blocklength=32 for left-to-right block decoding, or tau=None, stepsperblock=k for a fixed budget. The 50B-token checkpoint from the paper is…

Read IFML's full model card

A masked diffusion language model adapted from Qwen3.5-2B. The backbone is hybrid: only its attention layers are made bidirectional, and the Gated DeltaNet layers stay causal. This is a base model, with no instruction tuning.

Paper: dQwen3.5: Hybrid-Attention Diffusion Language Models. Code: https://github.com/AntonXue/dQwen

The dQwen3.5 family

model total params backbone
dQwen3.5-0.8B-Base 0.75B hybrid
dQwen3.5-2B-Base (this repo) 1.88B hybrid
dQwen3.5-4B-Base 4.21B hybrid
dQwen3.5-9B-Base 8.95B hybrid
dQwen3-1.7B-Base 1.72B full attention (control)

Quickstart

Needs a CUDA GPU and transformers>=5.13 (tested with torch 2.7.1+cu128, flash-linear-attention 0.5.1).

import torch
from transformers import AutoModel

model = AutoModel.from_pretrained("UT-IFML/dQwen3.5-2B-Base", trust_remote_code=True, dtype=torch.bfloat16).cuda().eval()
print(model.generate("def fibonacci(n):", gen_length=512, stop_strings=["\ndef "]).text)

generate decodes the whole canvas at once, committing positions above a confidence threshold (tau=0.9); pass block_length=32 for left-to-right block decoding, or tau=None, steps_per_block=k for a fixed budget. The 50B-token checkpoint from the paper is revision="step25000-swa".

Citation

@article{xue2026dqwen,
  title   = {dQwen3.5: Hybrid-Attention Diffusion Language Models},
  author  = {Xue, Anton and Rout, Litu and Akella, Aditya and Klivans, Adam and Sanghavi, Sujay and Shakkottai, Sanjay},
  journal = {arXiv preprint arXiv:2609.20751},
  year    = {2026}
}

Configuration

Architecture
DiffuQwen35
Context length (tokens)
262,144
Layers
24
Hidden size
2,048
Feed-forward size
6,144
Attention heads
8
Key/value heads
2
Head dimension
256
Vocabulary size
248,320
Model type
diffuqwen35

Identity and Version

Repository
UT-IFML/dQwen3.5-2B-Base
Publisher
IFML
Task
Text generation
Modality
Text
Library
transformers
Parameters
1.9B parameters
Languages
Not stated by the source
Revision
297e02d4a8f007301f45a175912c3252b509fe31
First published
2026-08-06
Last updated
2026-09-18

Files and Weights

11 files, 3.8 GB in total. The weights are 1 file totalling 3.8 GB in safetensors.

Weights1 file · 3.8 GB
Configuration4 files · 17.0 KB
Tokenizer2 files · 20.0 MB
Documentation1 file · 2.1 KB
Other2 files · 1.1 MB
Repository1 file · 1.6 KB
Every file
FileTypeSizeSHA-256
model.safetensorsWeights3.8 GB eca1b87a5972
argparse.jsonConfiguration866 B —
config.jsonConfiguration1.9 KB —
generation_config.jsonConfiguration179 B —
modeling_dqwen3_5.pyConfiguration14.0 KB —
README.mdDocumentation2.1 KB —
chat_template.jinjaOther7.8 KB —
training_log.jsonlOther1.1 MB —
.gitattributesRepository1.6 KB —
tokenizer.jsonTokenizer20.0 MB 06b9509352d2
tokenizer_config.jsonTokenizer1.2 KB —

License and Download

License
apache-2.0
Access
Open weights, no gate
Download size
3.8 GB
Download from IFML

Released by IFML through its official repository on Hugging Face. Read the license.

Built From

Memory Requirements

PrecisionWeights in memory
As published3.8 GB
16-bit3.8 GB
8-bit1.9 GB
4-bit0.9 GB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Questions About dQwen3.5-2B-Base

How much GPU memory does dQwen3.5-2B-Base need?

About 4.5 GB at 16-bit and 1.1 GB at 4-bit: the weights (1.9B parameters) plus a working margin. A long context needs more.

What is the cheapest GPU to run dQwen3.5-2B-Base on?

At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.

Can I use dQwen3.5-2B-Base commercially?

Yes. dQwen3.5-2B-Base is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

What is dQwen3.5-2B-Base's context length?

262,144 tokens, from the maximum position embeddings in its published configuration.

Similar Models

VNPen is MewBaka Studio's visual-novel model series. The writer edition is for script writing, de-AI rewriting, and generating example scenes from a mood brief. Output format is one script line per line: speaker：text, with narration written as 旁白：. The base Qwen/Qwen3.5-2B is multimodal. Its checkpoint carries 297 model.visual. tensors (a depth-24 / hidden-1024 / patch-16 ViT) and 15 mtp. tensors for multi-token prediction. This project is text-only. In transformers, AutoModelForCausalLM on a qwen35 config builds Qwen35ForCausalLM over a Qwen35TextConfig — so the vision weights were never loaded at any point: not for training, not for merging, not for saving. Verified on the published…

Open weights apache-2.0 1.9B parameters 262,144 tokens transformers
View model

A fast and efficient ~1.5B model optimized for CPU inference. The model was refactored with BitNet features and an updated tokenizer that includes new Routing, Tool call, and Robotics tags. Built on a redesigned DeepSeek R1 architecture with native ternary (BitNet-style) support and ready-to-run GGUF quantizations. - JiRack is a cloud-ready model that helps save money on cloud infrastructure. It can be used as an expert model in RAG deployments, with the ONNX JiRack Java server as an alternative. - Benefits high quality CPU inference TQ2 on Llama.cpp and Ollama via QAT - Robotcs, Routing, Coding, Multimedia, Advanced tool calling via CMSManhattan/JiRackPrecisionTokenizer - We are working to…

Open weights mit 1.8B parameters 131,072 tokens
View model

Model · Text generation

Qwen3-1.7B

Qwen

Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models. Built upon extensive training, Qwen3 delivers groundbreaking advancements in reasoning, instruction-following, agent capabilities, and multilingual support, with the following key features: - Uniquely support of seamless switching between thinking mode (for complex logical reasoning, math, and coding) and non-thinking mode (for efficient, general-purpose dialogue) within single model, ensuring optimal performance across various scenarios. - Significantly enhancement in its reasoning capabilities, surpassing previous QwQ (in thinking mode) and…

Open weights apache-2.0 2B parameters 40,960 tokens transformers
View model

Model · Text generation

TopologicalQwen

Convergent Intelligence

Topology-Aware Knowledge Distillation from Qwen3-30B-A3B → 1.7B TopologicalQwen is a 1.7B parameter model distilled from Qwen3-30B-A3B using Topological Knowledge Distillation (TKD) — a methodology that treats the teacher's output distribution over a concatenated token stream as a bounded variation (BV) function and decomposes knowledge transfer into three channels via the Mesh Fundamental Identity: 1. Smooth distillation (AC component) — Standard KL divergence over regions where the teacher's distribution varies continuously. This is what every other KD method does and stops at. 2. Jump corrections (D^j f) — Explicit correction terms at conceptual boundaries where the teacher's…

Open weights apache-2.0 2B parameters 40,960 tokens transformers
View model

Extended Reasoning Distillation from Qwen3-30B-A3B-Thinking → 1.7B The most downloaded model in the Convergent Intelligence portfolio. Qwen3-1.7B-Thinking-Distil captures extended deliberation patterns from the Qwen3-30B-A3B Thinking teacher — the variant that generates long-form reasoning chains before committing to an answer — and compresses them into a 1.7B student via supervised fine-tuning on the longwriter-6k dataset. The Thinking teacher produces the richest signal of the three teacher variants in the DistilQwen family (Instruct, Thinking, Coder). Where Instruct distillation captures clean instruction-following and Coder captures hierarchical decomposition, Thinking distillation…

Open weights apache-2.0 2B parameters 40,960 tokens transformers
View model

A 1.7B-parameter causal language model distilled from Qwen3-30B-A3B on 6,122 STEM chain-of-thought samples using discrepancy-informed knowledge distillation. The training objective emphasizes proof structure, detects reasoning pivot tokens through token-level divergence dynamics, smooths high-entropy student singularities before distillation, and monitors structural drift through discrepancy energy. Standard knowledge distillation treats all tokens uniformly. Even proof-weighted approaches typically apply a static multiplier over the entire derivation span. That helps, but it still misses the internal structure of reasoning: some regions are smooth procedural continuation, while others are…

Open weights apache-2.0 2B parameters 40,960 tokens transformers
View model