Model library / incoai
Qwen3.8-27B-DFlash2 is a block-diffusion drafter model for speculative decoding of the Qwen3.8-27B target model.
Parameters in total
Context length in tokens, as published
Weights at the released precision
Licence declared on the repository
Overview
The DFlash2 draft model is designed to run inside a speculative decoding server, generating blocks of draft tokens that are verified by the Qwen3.8-27B target model. It predicts whole token blocks, keeps top candidates per position, and uses a lightweight selector to produce a coherent output path, matching the target model's distribution.
The model has 1.9 billion parameters, uses the DFlash2DraftModel architecture, and supports a context length of 262 144 tokens. It is released under the Apache-2.0 licence and stored in bfloat16 format. Deployment requires integration with a speculative decoding engine such as SGLang or vLLM, and it operates without expert layers.
Specification
| Repository | incoai/Qwen3.8-27B-DFlash2 |
| Publisher | incoai |
| Published | 18 August 2026 |
| Architecture | DFlash2DraftModel |
| Model type | qwen3 |
| Parameters | 1,924,404,480 |
| Layers | 5 |
| Hidden size | 5,120 |
| Attention heads | 32, 8 key/value heads |
| Context length | 262,144 tokens |
| Vocabulary | 248,320 tokens |
| Weight format | bfloat16 |
| Licence | apache-2.0 |
| Base model | Qwen/Qwen3.8-27B |
Memory
Parameter count multiplied by the bytes each format uses. The key-value cache comes on top.
| Precision | Weights | Notes |
|---|---|---|
| bf16 | 3.6 GiB | Full precision as released |
| fp8 | 1.8 GiB | 8-bit, near-lossless on most models |
| int4 | 0.9 GiB | 4-bit, smallest footprint |
The cache costs 20 KiB per token at 16-bit, so the full 262,144-token context of one request needs about 5.0 GiB. Concurrency multiplies that number, not the weights.
That figure is an upper bound: this model uses sliding-window attention with a 2,048-token window on some layers, so long requests cache less than this.
Hardware
At the released precision, with 90% of the card given to the server and 2.5 GiB kept for the runtime. The context column is what is left for the cache on that setup.
| GPU | Cards needed | Free for cache | Context it holds |
|---|---|---|---|
| RTX 4090 24GB | 1 | 15.5 GiB | the full 262,144 tokens |
| RTX 5090 32GB | 1 | 22.7 GiB | the full 262,144 tokens |
| L40S 48GB | 1 | 37.1 GiB | the full 262,144 tokens |
| A100 80GB | 1 | 65.9 GiB | the full 262,144 tokens |
| H100 80GB | 1 | 65.9 GiB | the full 262,144 tokens |
| RTX PRO 6000 96GB | 1 | 80.3 GiB | the full 262,144 tokens |
| H200 141GB | 1 | 120.8 GiB | the full 262,144 tokens |
Serving
A vLLM launch line for the shape above. Check the model card for a runtime the publisher recommends.
vllm serve incoai/Qwen3.8-27B-DFlash2 \ --max-model-len 262144 \ --gpu-memory-utilization 0.90
Questions
The weights take about 4 GiB at the released precision. That fits on one RTX 4090 24GB, which leaves roughly 16 GiB for the key-value cache.
The repository declares apache-2.0. Read the licence text before commercial use: the name alone does not tell you what is allowed.
The configuration allows 262,144 tokens. Whether the whole window is usable depends on the memory left for the cache, which the hardware table works out per card.
This page is a reference, and Aiqre does not serve this model today. We deploy open-weight models on dedicated EU hardware on request, so if you want an OpenAI-compatible endpoint for it, email [email protected] with the model name and roughly what volume you expect.
Also in the library
MiniCPM5-2B is a 2-billion-parameter dense transformer aimed at on-device and resource-constrained deployments.
K2-Horizon-0.9B is a 1.1 billion-parameter dense decoder-only model with a 128 k token context window.
MiniCPM5-2B is a 2.5-billion-parameter dense Llama-style causal language model released by openbmb.
Manacá-1B is a 1.7 billion-parameter Llama-style decoder-only model pretrained for Brazilian Portuguese.
Hosting
Tell us the model and roughly what volume you expect. We reply with a price and an OpenAI-compatible endpoint, with your prompts never stored.