Roadmap and status

Where ModelSaar is today, what the next step is, and what is planned but not implemented. Nothing in the "planned" list is available now.

Now: public alpha

• Implemented and tested: OpenAI-compatible gateway (chat completions with streaming, models, embeddings, health), Ollama and generic OpenAI-compatible adapters, three-state capability evidence with probes and fingerprints, deterministic routing with profiles and aliases, session and tool-call continuity, admission control with memory budgets, decision records and a status page, the CLI, client configuration generators, an evaluation harness with sandboxed scoring, and a fault-injection test suite (208 tests at the pinned revision, including the independent reviewer's 29 contract checks kept verbatim; the 6 sandbox tests need an isolation backend and are skipped where none exists).
• Exercised live: discovery, probes, routing, streaming, tool calls, tool-result continuity, session pinning and a three-model evaluation against Ollama 0.32.13 on an ASUS GX10 (Benchmarks).
• Review status (precise): an independent review of the first public revision (1122e1f) produced 18 grouped findings, addressed on the hardening branch with regression tests (F01–F18). A second independent review of that branch (80276bb) reproduced 11 further contract failures and 6 launcher/installer observations (R1–R8; occupancy after stream/probe timeouts, resident-memory accounting, embedding identity anchoring, stream validation, the beginner path, evaluator isolation/scoring/resume, publication guards). Those were corrected on 2026-10-12 and the reviewer's own tests pass and are part of the suite; F04, F08, F10, F13, F14 and F17 were reopened and re-closed by that pass. A third independent review of the corrected revision (2026-10-12) reproduced five further edge conditions (failed residency query, refused-connection residency, stream terminal ordering, evaluator probe admission, checkpoint alternatives); they were closed in the alpha pass and the reviewer's tests are in the suite. Independent verification of this exact candidate is pending. The installer pins the candidate.
• Not yet: a tagged release, a PyPI package or a container image; real-client workflow tests (Open WebUI, OpenClaw, OpenCode, Hermes); a real vLLM/llama.cpp run; real-model embeddings; a thinking-matched evaluation budget.

Next

1. Independent verification of the corrected hardening revision (the 2026-10-12 pass), then merge and a first tagged release (0.1.x) with a PyPI package and container image.
2. One real client workflow recorded end to end (Open WebUI is the quickest), then the others.
3. An evaluation with a reasoning-matched budget and all eight GX10 models, and a run against a generic OpenAI-compatible server.

Planned, not implemented

• Native decision-model adapters as routing classifiers and a decisions-style endpoint.
• Audio and video inputs; multimodal embeddings; OCR and guardrail pipelines as separate results.
• Responses API and Anthropic Messages with fail-closed translation.
• Exact tokenisation for more runtimes; better served-context detection for Ollama.
• Managed model lifecycle (keep-alive/unload) under leases and locks; composition with llama-swap.
• Cascades and ensembles; scheduling across hosts.
• Opt-in adaptive learning from outcomes with versioning, holdout evaluation, privacy controls and rollback.
• Optional LLM-judge scorers in the evaluation harness, with identity, bias and cost disclosed.

The repository's roadmap file is the source for this list; the changelog records what actually landed, and the post-alpha backlog lists what the alpha deliberately leaves unverified or disabled (macOS code-execution sandbox, real-client walkthroughs, a live run under the current scorer).