DEV.co
Open-Source LLM · z-lab

Qwen3.6-35B-A3B-DFlash

Qwen3.6-35B-A3B-DFlash is a specialized 35B-parameter draft model designed for speculative decoding—it is not a standalone LLM. It works alongside Qwen3.6-35B-A3B to propose multiple tokens in parallel using block diffusion, which the target model then verifies. This approach aims to accelerate inference throughput (up to 3.61x speedup at concurrency 1) without changing the output distribution. It requires SGLang or vLLM (draft PR) and is licensed under Apache 2.0.

Source: HuggingFace — huggingface.co/z-lab/Qwen3.6-35B-A3B-DFlash
386M
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
182.9k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerz-lab
Parameters386M
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads182.9k
Likes260
Last updated2026-06-19
Sourcez-lab/Qwen3.6-35B-A3B-DFlash

What Qwen3.6-35B-A3B-DFlash is

A 35B diffusion-based draft model trained by Z-Lab and Modal on 40k sequence length with sliding-window attention. Deployed as part of a speculative decoding pipeline where it generates token proposals that are verified by the target Qwen3.6-35B-A3B model. Benchmarks show 2.89x–3.61x throughput gains across GSM8K, MATH500, HumanEval, MBPP, and MT-Bench at varying concurrency levels on 1x NVIDIA B200 GPU.

Quickstart

Run Qwen3.6-35B-A3B-DFlash locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="z-lab/Qwen3.6-35B-A3B-DFlash")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

High-Throughput Inference on Constrained GPU Resources

Deploy DFlash with target model on a single B200/H100 to achieve 2–3.6x throughput gains without retraining. Ideal when latency per token is acceptable but total request throughput is the bottleneck.

Batch/Long-Context Processing Workloads

Use DFlash at concurrency 32+ to balance accept length with batch efficiency. Benchmarks show consistent 2–2.5x speedups at high concurrency, suitable for document QA, summarization, and reasoning tasks (GSM8K, MATH).

Research/Experimentation on Speculative Decoding Algorithms

Integrate with SGLang or vLLM to prototype and benchmark diffusion-based draft models. Published paper (arxiv:2602.06036) and open-source code facilitate reproducibility and extension.

Running & fine-tuning it

ESTIMATE: ~85–90 GB total VRAM for both draft (DFlash, ~35B params, bfloat16 ~70 GB) + target (Qwen3.6-35B-A3B, ~35B params, bfloat16 ~70 GB) on a single GPU. Benchmarks run on 1x NVIDIA B200; H100 (80GB) will fit with tensor parallelism (tp-size > 1) or offloading. Requires CUDA 12.0+, recent SGLang build, and FlashAttention/TRT-LLM backends for optimal performance.

DFlash is a pretrained draft model and Not explicitly positioned for fine-tuning. It is a diffusion model coupled to a specific target model architecture. Fine-tuning or LoRA adaptation is not documented. To speculate on other domains, retrain from scratch or use the base Qwen3.6-35B-A3B; modifying DFlash alone may break speculative decoding alignment. Requires review with Z-Lab or Modal for customization guidance.

When to avoid it — and what to weigh

  • You Need a Standalone Language Model — DFlash is a draft model only; it cannot generate text independently. You must pair it with the full Qwen3.6-35B-A3B target model, doubling memory overhead and complexity.
  • You Require Sub-10ms Token Latency — Speculative decoding introduces verification overhead. Single-token latency may not improve or may increase compared to direct inference; prioritize throughput-optimized serving.
  • You Need vLLM as Your Sole Inference Backend — vLLM support is in draft PR status (vllm-project/vllm#40898) and not yet merged. SGLang is the primary validated backend; vLLM support is not production-ready.
  • Your Target Model is Not Qwen3.6-35B-A3B — This model is tuned specifically for Qwen3.6-35B-A3B. Pairing with other target models is unsupported and accept lengths / speedup may degrade significantly.

License & commercial use

Apache License 2.0 (SPDX: apache-2.0). This is a permissive OSI-approved open-source license.

Apache 2.0 permits commercial use, modification, and distribution under the same license with attribution. No restrictions on commercial deployment stated in the license. However, verify with Z-Lab and Modal that use of their jointly trained model aligns with their terms of service and any supporting compute/cloud agreements. Requires review if you plan to resell or embed in a commercial product without modification.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityHigh
DEV.co fitGood
Assessment confidenceHigh
Security considerations

DFlash inherits the security posture of Qwen3.6-35B-A3B and the SGLang/vLLM inference servers. Custom code is enabled (tag: custom_code). Verify that SGLang and vLLM builds are up-to-date and sourced from official channels. Run DFlash in isolated environments or VPCs; monitor GPU memory and CPU/network usage for unexpected load. No formal security audit or threat model is stated. Treat similarly to any large model serving infrastructure.

Alternatives to consider

Qwen's Built-in MTP Draft Path

Native multi-token prediction within Qwen3.6-35B-A3B. Avoids dual-model overhead but typically achieves lower speedups (1.5–2.2x per benchmarks) and worse accept lengths than DFlash block=8/16.

Medusa (Speculative Decoding)

Another draft model approach that grows a lightweight tree of predictions. Simpler training than diffusion, but requires Medusa-specific checkpoints and integration. May not outperform DFlash on this target.

Llama 3.1-70B or Mistral-Large with Batching

Larger standalone models with native throughput via batching. Avoid dual-model complexity but require more VRAM; benchmarks not directly comparable. Consider if you need a general-purpose 70B+ model.

Software development agency

Ship Qwen3.6-35B-A3B-DFlash with senior software developers

Integrate DFlash with SGLang to accelerate Qwen inference on your GPU infrastructure. Review our deployment guide, benchmark your target workloads, and verify hardware requirements with our team.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

Qwen3.6-35B-A3B-DFlash FAQ

Can I use DFlash for commercial inference without contacting Z-Lab?
Apache 2.0 permits commercial use. However, since Z-Lab and Modal jointly trained and maintain this model, review their websites and terms of service for any additional restrictions or support requirements before deploying in production.
What GPU do I need to run DFlash + target model?
Benchmarks use 1x NVIDIA B200 (192 GB HBM). For smaller GPUs (H100 80 GB, A100 80 GB), enable tensor parallelism (tp-size 2+) or use quantization. With tp-size=1, you will likely need two GPUs or aggressive offloading.
Does DFlash work with other target models besides Qwen3.6-35B-A3B?
No. DFlash is trained specifically to speculate for Qwen3.6-35B-A3B. Using it with other models will degrade performance and is not supported.
Is vLLM support ready for production?
No. vLLM support is in draft PR status (vllm-project/vllm#40898) and not yet merged. SGLang is the recommended and tested backend. Revisit vLLM in future releases.

Work with a software development agency

DEV.co helps companies turn open-source tools like Qwen3.6-35B-A3B-DFlash into production software. Our software development services cover the full lifecycle — architecture, web development, integration, and maintenance — delivered by software developers and web developers who ship. Engage our software development agency to implement or customize it for your open-source llms stack.

Ready to Deploy High-Throughput LLM Inference?

Integrate DFlash with SGLang to accelerate Qwen inference on your GPU infrastructure. Review our deployment guide, benchmark your target workloads, and verify hardware requirements with our team.