Skip to content

Releases: rmacy/qwen36-b70-mtp2

Release list

Qwen3.6-27B B70 MTP2 v0.1.3 — audited release

Qwen3.6-27B B70 MTP2 v0.1.3 — audited release Latest
Latest

Choose a tag to compare

View all tags
@rmacy rmacy released this 14 Aug 01:57
v0.1.3
db10937

Audited release of the bounded-MTP2 Qwen3.6-27B FP8 runtime for 2x Intel Arc Pro B70.

Images (identical digest):

  • ghcr.io/rmacy/qwen36-b70-mtp2:0.1.3
  • us-central1-docker.pkg.dev/home-504803/open-models/qwen36-b70-mtp2:0.1.3
  • sha256:a77db6582579e417f07e3944ab5092f8989bc42eb7d40fb72d4b828d7e0bd3f0

Validated profile:

  • 52.94 tok/s C1 median versus 31.90 tok/s without MTP
  • 169.48–175.94 tok/s cold aggregate at C4
  • 267.37–282.57 tok/s cold aggregate at C8
  • quality tied the separately hosted copy on the frozen score and all 82 blinded comparisons
  • 6,981/6,981 correct requests in a 3,604-second mixed soak

Release hardening:

  • final-filesystem MTP, launcher, API-key, and CVE-2026-48746 exploit regression tests
  • cache-resistant benchmark smoke test with the nonce in the first prompt block
  • optional API_KEY support through both launcher paths
  • commit-pinned GitHub Actions and OCI-aware anonymous-pull validation
  • Ubuntu linux-libc-dev pinned to the fixed 6.8.0-137.137 package

The exact published image was pulled anonymously and passed every embedded runtime test. Its five performance-critical vLLM files are byte-identical to the validated production runtime. Trivy 0.73.0 found zero secrets. Its two critical records are duplicate version-based matches for CVE-2026-48746; the affected path is backported and regression-tested. The image contains no model weights or private deployment data.

Assets 2
Loading

v0.1.0 — validated dual-BMG MTP2 runtime

v0.1.0 — validated dual-BMG MTP2 runtime

Choose a tag to compare

View all tags
@rmacy rmacy released this 14 Aug 00:02
v0.1.0
30e2bd8

First sanitized public release of the Qwen3.6-27B FP8 MTP2 reproduction kit for dual Intel BMG / Arc Pro B70 GPUs.

Published containers:

  • ghcr.io/rmacy/qwen36-b70-mtp2:0.1.0
  • us-central1-docker.pkg.dev/home-504803/open-models/qwen36-b70-mtp2:0.1.0
  • Digest: sha256:8593d10b6217957da17e336c81ea61029f1e402fed19a6bea6e5605240983173

Validated results:

  • C1 decode median: 52.94 tok/s
  • Cold C4 aggregate: 169.48–175.94 tok/s
  • Cold C8 aggregate: 267.37–282.57 tok/s
  • Mixed soak: 6,981/6,981 successful requests

Security:

  • No model checkpoint, credentials, private endpoints, personal paths, or benchmark responses are included.
  • The image backports the upstream CVE-2026-48746 authentication fix and includes a regression test.
  • Read SECURITY.md before exposing the OpenAI-compatible endpoint.
Loading