Muse Glimmer B70 v0.1.5

Fixes first-start full-context initialization on an empty compile cache by changing the published GPU reservation default from 0.72 to 0.74. Model precision, TP=2, 131,072-token context, 4,096-token batch ceiling, DFlash15, XPU graph mode, and bounded vision are unchanged.

Published images (identical digest):

• ghcr.io/rmacy/glimmer-b70-vllm:0.1.5
• us-central1-docker.pkg.dev/home-504803/open-models/glimmer-b70-vllm:0.1.5
• sha256:8fc1eb0d459a4602e39e0af6551858e667ce9856c0e5a8db41970c732eab5617

Acceptance on the exact public digest with a new cache: ready in 180 seconds, 148,010 KV tokens at full context, deterministic text pass, streaming and nonstreaming ATEM tool-call passes, tool-result continuation pass, and 1,792-by-1,792 maximum-image pass. The image contains no model weights or private configuration.

Trivy 0.73.0: zero secrets; the two critical records are duplicate version matches for CVE-2026-48746, whose upstream fix is backported and covered by the included regression test. See SECURITY.md for the inherited dependency baseline.

Muse Glimmer B70 v0.1.4 — audited FP8 release

Audited release of the full-parameter FP8 Muse Glimmer 30B runtime for 2x Intel Arc Pro B70.

Images (identical digest):

• ghcr.io/rmacy/glimmer-b70-vllm:0.1.4
• us-central1-docker.pkg.dev/home-504803/open-models/glimmer-b70-vllm:0.1.4
• sha256:7970aadf82197e69a6abb648cb615c0771b89ecb73405a4f4dec705cbd106ed9

Validated profile:

• 88.43 tok/s median at approximately 32K context
• 84.84 tok/s median at approximately 64K context
• 79.59 tok/s across the 12-request approximately 126K exact-retrieval soak
• 131,072-token context, official DFlash, ATEM tools, and bounded-memory vision

Release hardening:

• every patch/auth/parser/launcher test now runs against the final image filesystem
• exact measured Hugging Face revisions are pinned in the setup instructions
• Compose forwards API_KEY and defaults to this release tag
• GitHub Actions dependencies are commit-pinned and the workflow verifies an anonymous pull
• Ubuntu linux-libc-dev is pinned to the fixed 6.8.0-137.137 package

The exact published image was pulled anonymously and passed all runtime, DFlash, vision-parity, parser, launcher, and CVE-2026-48746 exploit regression checks. 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.

Muse Glimmer B70 v0.1.2

Hardened, reproducible Muse Glimmer 30B runtime for two Intel Arc Pro B70 GPUs. This release corrects the checked-in Compose default so a plain GHCR_OWNER=rmacy docker compose up -d selects the hardened image.

Published images:

• ghcr.io/rmacy/glimmer-b70-vllm:0.1.2
• us-central1-docker.pkg.dev/home-504803/open-models/glimmer-b70-vllm:0.1.2
• Immutable digest: sha256:3ddf64c04ed0250a84b413cda63c24222523b5ddc73f8c1ed8a8b73b17b98b23

Validated profile:

• Full-parameter FP8 target plus official FP8 DFlash assistant
• Native 131,072-token context, tools, and bounded-memory vision
• 88.43 tok/s median at 32,023 input tokens
• 84.84 tok/s median at 63,848 input tokens
• 79.25 tok/s median at 125,991 input tokens
• 12/12 ~126K soak requests, 79.59 tok/s median

Security and equivalence:

• Includes the upstream CVE-2026-48746 authentication-bypass fix and loopback-only Compose publishing
• Authentication regression passed on the published image
• The complete allowlisted Docker build context is byte-identical to v0.1.1; only README, SECURITY.md, and Compose metadata changed
• Both registries were verified anonymously readable

See SECURITY.md for the pinned Intel base image advisory inventory.

Muse Glimmer B70 runtime 0.1.1

Security-aligned release of the native Muse Glimmer 30B runtime for two Intel Arc Pro B70 GPUs.

Published containers:

• ghcr.io/rmacy/glimmer-b70-vllm:0.1.1
• us-central1-docker.pkg.dev/home-504803/open-models/glimmer-b70-vllm:0.1.1
• Digest: sha256:cf6c3513115d45473136b504b8102720d4c098e2726ac8841aef9d9be5ed4d41

Measured release profile:

• 88.43 tok/s median at approximately 32K context
• 84.84 tok/s median at approximately 64K context
• 79.25 tok/s median at approximately 126K context
• 12/12 exact-retrieval soak at 79.59 tok/s median

Changes since 0.1.0:

• Backports the upstream CVE-2026-48746 vLLM authentication fix.
• Adds ordinary and malicious-Host authentication regression tests.
• Binds the supplied Compose profile to loopback by default.
• Adds optional API_KEY bearer-token authentication.
• Adds OCI source/license metadata and SECURITY.md.

All nine Glimmer/DFlash/XPU performance-critical files match 0.1.0 byte-for-byte. Source, image configuration/history, and image filesystem scans found no credentials or private deployment data. The inherited base still version-scans with 18 critical and 228 high findings; the critical set is documented in SECURITY.md, including the backported vLLM finding.

Muse Glimmer B70 runtime 0.1.0

Native Intel-vLLM runtime for full-parameter FP8 Meta Muse Glimmer 30B on two Intel Arc Pro B70 GPUs. Includes official DFlash support, 131,072-token context, ATEM tools, bounded-memory vision, a self-contained launcher/parser, and the measured production defaults. Model weights are not included. Pull: docker pull ghcr.io/rmacy/glimmer-b70-vllm:0.1.0