OPEN LINK ↗
// 1h agoNEWS
Qwen4-27B Leaks Through Intel Kernel PR
A merged Intel vLLM XPU kernel PR appears to reveal that Alibaba’s unreleased Qwen4-27B is a dense model optimized for FP8 on two cards and INT4 on one or two. The code also references four-token multi-token prediction, hyper-connections, and sparse attention, though no official weights or model card exist.
// ANALYSIS
This is a compelling architecture leak, but still an inference from implementation details rather than an Alibaba specification.
- –Dense 27B inference could make Qwen4 substantially more accessible than a large MoE deployment.
- –INT4 support across one or two GPUs points toward practical local serving on consumer hardware.
- –Four-token MTP could materially improve decoding speed if the draft head performs well.
- –The PR suggests four residual streams, gated recurrent-style components, and QSA attention, but several details remain unconfirmed.
- –Reported benchmarks are preliminary, and QSA could not be fully tested without the real checkpoint.
// TAGS
qwen4-27bllmopen-weightsquantizationinferencegpuopen-source
DISCOVERED
1h ago
2026-10-25
PUBLISHED
1h ago
2026-10-25
RELEVANCE
9/ 10
AUTHOR
0x0SojalSec