The honest list — items in 5.0 that are either in a preview state, currently blocked behind a fix, or queued for an upcoming 5.0.x patch. Updated as items resolve. The intent is transparency, not exhaustiveness: every entry here is something a customer could plausibly hit and want context on.
Items caught during the v5.0.0 GA close. These are the 5.0.1-alpha.1 work-items the platform team picks up in the first week. Customer-facing impact ranges from minor (display freshness) to opt-in workarounds (CLI environment, demo connectors).
The ai-papers-at knowledge base imported from the 4.x corpus is thinner than the historical 4.x knowledge base. Specific topics — Hochreiter-authored papers in particular — are sparsely represented in the GA cut. Re-import is in flight; 5.0.1 ships with the rebuilt corpus.
The most-used and recently-used badges in the chat model picker need one warm-up pass to populate after install. The chat shell handles this automatically as customers use models; admins who want pre-populated stats on day one can request the warm-up routine via support@eldric.ai.
The namespace count shown next to a knowledge base in the Admin Console can be stale immediately after upload. Document and chunk counts are correct; the namespace count refreshes on the next aggregate cycle. Fix lands in 5.0.1.
The demo ClickHouse and Postgres connectors aren't registered with the platform by default in 5.0 GA. Customers who want them available run the setup scripts in the operations guide. 5.0.1 wires them as default-enabled with a one-click toggle in the Admin Console.
Locally-built command-line clients (built from Enterprise-tier source) need ELDRIC_EDGE_URL and ELDRIC_API_KEY set in the environment to reach a cluster. RPM-installed CLI reads configuration from /etc/eldric/cli.conf automatically and needs no setup.
The per-tenant theme and branding endpoints — /api/v1/tenants/{id}/theme and /api/v1/tenants/{id}/branding/logo — aren't yet reachable through the public Edge gateway. Pro+ customers configuring custom branding call them from inside the cluster network on the 5.0 line. The Edge allowlist update lands in 5.0.1.
Pure chat sustains ~793 req/s at 32 concurrent connections with 41 ms median latency on the reference cluster — comfortable. Knowledge-base search at four concurrent connections still hits a ~7-second p50 latency cliff. We publish the number and we're actively working on it; the EMM compressed-retrieval preview (per-knowledge-base opt-in) is the path that closes most of this gap for 5.0 customers who can use it. Full fix lands in an upcoming 5.0.x patch.
The OPC-UA closed-loop transport's C++ source ships in 5.0, but the open62541 library isn't yet bundled into the released Fedora 43 RPMs — customers running stock release RPMs find the binary doesn't link the OPC-UA symbols. The transport stays marked Preview on xLSTM & IoT until the next RPM cut adds the dependency. WebSocket, Modbus and MQTT-publish are GA and unaffected.
The publish half (sending policy outputs to actuator topics) is GA. The subscribe half (pulling observations from sensor topics through the policy worker) stays preview through the 5.0 line; the per-binding background subscriber + ring buffer lands in an upcoming 5.0.x patch.
Fedora / RHEL / CentOS / Rocky / Alma RPMs ship at GA. Ubuntu 24.04 and Debian 12 native DEBs follow shortly after GA. Customers on Ubuntu / Debian today can run 5.0 via WSL2 or container — install instructions live on the Ubuntu 24.04 and Debian 12 pages.
Some models can use tools — web search, real-time data, file access, multi-agent task planning — and some can't. If you pick a model without tool support and ask a question that needs tools, Eldric now handles it for you: it either automatically switches you to a tool-capable model and tells you which one answered, or asks you to pick a different model and shows which ones support tools for your question. This works the same whether you use the web chat, the iOS app, the command-line client, or the API directly — so questions that need to fetch external data no longer come back generic, empty, or with invented citations. If you'd rather choose yourself, tool-capable models include llama3.2:3b, qwen2.5:7b, and the Claude and GPT-4 families.
The Bosch IoT plugin installs from Admin → Marketplace: search for "Bosch", click Install. If the install button shows "available" for a plugin that's already installed, refresh the page after about 5 minutes — the marketplace state-sync converges shortly. A faster sync lands in the next 5.0 alpha.
Two admin tabs in Eldric 5.0 — xLSTM and NOVA — are placeholders for subsystems shipping in an upcoming 5.0.x patch. On a 5.0 install you'll see the tabs but the underlying daemons aren't part of the 5.0 binary. The §42 Autonomous Dreaming cycle is in 5.0 and lives at Cluster Ops → Dreams (not the NOVA tab). The full NOVA subsystem (goals, self-modification, safety constraints) and the xLSTM workload daemon (policy / forecast / encode / retrieve) land in an upcoming 5.0.x patch — see What's Next.
These are working features in 5.0 that customers can turn on but that haven't yet finished hardening. They have customer-facing docs and are supported on Pro+ tiers as preview.
The inference daemon consults the matrix-memory layer directly at prompt boundaries. Sub-2 ms per-token overhead on CPU. Pro+ opt-in per request. See smart memory inference.
EMM Compressed variant of the matrix-memory layer — smaller .emm files, faster at high concurrency, small accuracy trade-off on edge-case queries. Per-knowledge-base opt-in. See advanced retrieval.
Single-pass neural classifier trained on cluster routing data; lower-latency routing decisions; the LLM stays in the loop for ambiguous queries. Pro+ opt-in. Training is underway — the GQA-distilled checkpoint publishes shortly after GA.
Stage Manager + Split View are supported. Floating windows above other apps are an upcoming 5.0.x target — the UISceneDelegate work is in flight but the multi-tenant scene routing isn't fully exercised yet.
Items that are scoped, designed, but not part of 5.0 GA. The 5.0 release ships when its surface is right, not when every wishlist item is in.
5.0 ships four built-in roles — Viewer / Developer / Admin / SuperAdmin. Customer-defined role names and finer-grained capability bundles arrive in an upcoming 5.0.x patch.
A 5.0 cluster is a single-region installation. The federation layer for stitching two regional clusters into a single logical platform — with replicated knowledge bundles and cross-region routing — lands in an upcoming 5.0.x patch. Today, the .nexus bundle export is the lightweight path.
The experimental self-improvement module remains explicitly experimental and off by default. Production-status (and customer-facing docs beyond the research framing) comes later in 5.0.x.
The Level-1 native xLSTM backend (JAX-XLA + IREE) compiles cleanly in 5.0 with the build flag disabled. Wiring it through as the default backend on Apple Silicon lands in an upcoming 5.0.x patch once we've completed the IREE-link integration.
The EMM Compressed retrieval format ships at full precision in 5.0. The int8-quantized variant — for even smaller .emm files on edge nodes — lands in an upcoming 5.0.x patch.
5.0 ships five themes (Frost + four others) with cluster sync. Customer-customisable colour palettes inside the theme editor arrive in an upcoming 5.0.x patch.
The Apple platform codebase is positioned to add a visionOS build later, but no visionOS target ships at 5.0 GA. No timeline commitment.
X-Foo header was effectively invisible to the gateway because of HTTP/2 lowercasing; fixed across the gateway path.Free tier: open an issue describing what you saw, what you expected, and the steps to reproduce. Paid-tier customers: mail support@eldric.ai with your license ID — prioritised handling.
For install-time problems specifically, install troubleshooting covers the nine common failure modes (GPG fail, port conflicts, admin conflicts, Gatekeeper, license signature, etc.) with recoveries.