DFlash and DDTree run Qwen3.6-27B on one RTX 3090
A post dated April 23, 2026 reports 73 tokens per second with speculative decoding on a single RTX 3090 and identifies architectural compatibility as a factor in running the model.
Source: DFlash e DDTree executam Qwen3.6-27B em uma RTX 3090 (x.com). Text prepared with AI from this source.
What happened and what to do
In a post published on April 23, 2026, the author reports running Qwen3.6-27B at 73 tokens per second on a single RTX 3090, using speculative decoding with DFlash and DDTree. According to the post, execution is possible because its architecture and layer and head dimensions match those of Qwen3.5, although throughput is lower. The case shows how architectural compatibility can enable the technique before dedicated upstream support is available.
A company can test this approach with its own models and workloads, measuring latency, throughput, memory use, and stability on available GPUs. An implementation could include a private inference environment, reproducible benchmarks, and monitoring to compare the technique with conventional execution and determine whether it meets operational requirements.
How the consultancy can help
Wendelmaques can assess model and infrastructure compatibility, define a testing scope, and implement and operate an inference and monitoring environment suited to the case.
Next step
Send a short description of your case to receive a scoped proposal.
Let's assess your project
I reply with scope and a proposal.