Qwen4-27B Hasn't Been Released Yet, but a Kernel PR Has Already Leaked the Configuration
Intel's vLLM XPU kernel merge records show configurations such as Qwen4-27B dense, INT4 on 1 to 2 cards, FP8 TP2, and MTP draft=4. The community is both guessing the architecture and calculating whether 24GB GPUs can run it. There is no official confirmation yet.
Qwen4-27B hasn't been released yet, but Intel's vLLM XPU kernel merge records have revealed its shape first. The X account @0x0SojalSec says someone reverse-engineered architecture clues for Qwen4-27B from a kernel PR that has already been merged: dense 27B, INT4 on 1 to 2 cards, MTP with 4 draft tokens, FP8 TP2, 4 residual streams/hyper-connections, and a QSA sparse attention kernel.
The commit information in the screenshot is more specific. The committer hzjane merged 10 commits into analytics-zoo:downstream/v0.31.0, indicating that these SYCL/SYCL-TLA kernels serve downstream v0.31.0's Qwen3.8-Flash-Next TP4 INT4, TP8 INT4/FP8, as well as Qwen4-27B dense TP2 FP8 and TP1/TP2 INT4. They cover M=1-8, MTP draft=4/B1, and large prefill, with the goal of reducing wasted compute, memory round trips, and host submission overhead. The accompanying model, metadata, scheduler, and FULL graph integration are in vLLM#51.
Some commenters noted that these are SYCL kernels, so the 1-to-2-card INT4 path points to Intel XPU. The PR only lists coverage, not speed or INT4 quality numbers. Guessing architecture from a kernel PR was compared by one commenter to reading tea leaves, except the commit message is more detailed. The metaphor is pretty apt: kernel shapes can reveal deployment targets, but they are still far from a complete architecture and real-world experience.
Community reactions split into two camps. One camp is doing VRAM math: one commenter estimates that for a 27B dense model with INT4, the weights are about 13.5GB, not counting KV cache. It can run on 1 to 2 cards, but phone-class memory is out of the question; it lands on desktop first. The other camp is correcting things: 27B is already dense, not MoE, and the "MoE cluster" line in the original tweet was mocked by commenters as sounding bot-written. Some also worry that its reasoning is too cautious and hope the MTP draft head is more confident, rather than agonizing over simple changes like 3.8 does.
There is no official confirmation of Qwen4-27B's existence or specs. What can be confirmed for now is that these shapes appeared in a merged Intel vLLM XPU kernel PR. For local deployment, INT4 plus 1 to 2 cards is a line worth watching; for architecture speculation, don't mistake tea leaves for a model card just yet.
发布时间: 2026-10-25 20:41