Bielik-11B-v3.1-DFlash

Bielik-11B-v3.1-DFlash

Serves as the DFlash draft model that pairs with Bielik-11B-v3.1-Instruct for speculative decoding in SGLang or vLLM. It is not a standalone language model. For people who host Bielik.

Made by speakleash on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Bielik-11B-v3.1-DFlash is a small draft model, with 1.2B parameters, built by SpeakLeash and ACK Cyfronet AGH. It works only next to its target model, Bielik-11B-v3.1-Instruct, in a speculative decoding setup run by an inference server. It was trained on the same data as the target and is not a standalone language model.

Who it suits: It suits people who already serve Bielik-11B-v3.1-Instruct and want to add DFlash speculative decoding to that server.

Industry | Who uses it | How they use it | Example
AI infrastructure | Teams hosting Bielik-11B-v3.1-Instruct | Add the draft model to their SGLang or vLLM launch command | A team adds the speculative flags to its existing server command
Polish-language software | Developers building apps for Polish users | Serve the target model with the draft model attached | A developer tests the paired setup before shipping a chat feature
Research | People studying speculative decoding | Try DFlash drafting with a Polish-focused model | A researcher compares runs with and without the draft model

Anyone who wants a model to chat with directly can skip it, because the draft model cannot answer on its own.

Where to use it: It runs inside an inference server that supports DFlash, and the files download on Windows, macOS or Linux with Python.

• SGLang, in a version above 0.5.6: launch the server with the DFLASH algorithm
• vLLM, in a version above 0.20.0: serve the target model with a speculative config
• Windows, macOS or Linux: download the files with the Hugging Face tools

Use case | Who | What they get back
Host Bielik with drafting | Platform teams | A server that runs the target model with the draft model attached
Polish and European-language apps | Developers | Replies from Bielik-11B-v3.1-Instruct served through the paired setup
Test speculative decoding | Researchers | A working DFlash configuration to measure

How to set it up: You need a few things first.

• Python and the huggingface_hub package for the download
• SGLang above version 0.5.6, or vLLM above version 0.20.0
• A running copy of Bielik-11B-v3.1-Instruct, the model this draft is paired with
• About 2.3 GB of disk for the draft model in BF16

1. Install the Hugging Face hub package.
2. Download the draft model files if you want a local copy.
3. Install SGLang or vLLM in the version the card lists.
4. Launch the server with the target model and the draft model, as shown under How to use it.

pip install -U huggingface_hub
hf download speakleash/Bielik-11B-v3.1-DFlash

How to use it: Run it behind a server such as SGLang or vLLM, whichever has DFlash drafting built in. The card gives one command for each.

export SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1

python -m sglang.launch_server \
    --model-path speakleash/Bielik-11B-v3.1-Instruct \
    --speculative-algorithm DFLASH \
    --speculative-draft-model-path speakleash/Bielik-11B-v3.1-DFlash \
    --speculative-num-draft-tokens 16

vllm serve speakleash/Bielik-11B-v3.1-Instruct \
  --speculative-config '{"method": "dflash", "model": "speakleash/Bielik-11B-v3.1-DFlash", "num_speculative_tokens": 15}' \
  --attention-backend flash_attn \
  --max-num-batched-tokens 32768

Tool | What it does
SGLang launch_server | Runs the target model with the DFLASH algorithm and the draft model path
vllm serve | Serves the target model with a speculative config that names dflash and the draft model
SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN | An environment variable the card sets to 1 before it starts SGLang
--speculative-num-draft-tokens | The draft token count, 16 in the SGLang command, while the vLLM config uses 15

Example: You start the vLLM server with the command above and send the target model a chat request, such as "Summarize this paragraph in Polish". The reply is written by Bielik-11B-v3.1-Instruct. The draft model sits behind it through the speculative config.

For newcomers and experts: This file does nothing by itself. It is a helper for another model, Bielik-11B-v3.1-Instruct, and only makes sense once that model runs on a server. Your first small step is to start the SGLang or vLLM command from the card with both model names filled in.

The draft model was trained on the target's dataset, with a custom implementation on top of the ALLaMo framework. The card's commands use 16 draft tokens in SGLang and 15 in vLLM. The trade-off is compatibility, since you need a server version that supports DFlash and a target model to pair it with.

Start the server with Bielik-11B-v3.1-Instruct
Attach the DFlash draft model through the flags
The draft model proposes tokens
The target model produces the reply

Good to know:

• The card lists 32 European languages, optimized for Polish, while the Hugging Face record shows Polish and English.
• SGLang must be newer than 0.5.6 and vLLM newer than 0.20.0.
• The card says its development and training were supported by a computational grant and ran on the Helios supercomputer.
• License: Apache 2.0, as the card and the Hugging Face record both list it.
• Cost to run: none named, no API key, account or paid service is mentioned.
• Data it reaches: it downloads files from Hugging Face and runs inside your own inference server.
• Not for: use as a standalone language model.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 1.2B
• Weights: about 2.3 GB in BF16; about 587 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 36
• Likes: 3
• License: apache-2.0
• Languages: pl, en
• Quantized versions: 2
• Tags: transformers, safetensors, qwen3, dflash, speculative-decoding, block-diffusion, draft-model, efficiency, bielik, diffusion-language-model, text-generation, multilingual

Links

• Model: https://huggingface.co/speakleash/Bielik-11B-v3.1-DFlash
• Files: https://huggingface.co/speakleash/Bielik-11B-v3.1-DFlash/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download speakleash/Bielik-11B-v3.1-DFlash

Then follow the usage in its model card. Nothing in this kit runs by itself.