Portable across local models.
Memstead's authoring skills — ingest, sync, verify — run through the Model Context Protocol, so they work against any model that calls tools, including ones on your own machine. No cloud. No API key. File contents the agent reads never leave the host.
This page is a proof, not a promise: every "verified" mark below traces to a real run from local-ai/. Where a model hasn't been put through tool-calling yet, it's shown as a candidate — not asserted.
The models
Three 4-bit local models, run via an oMLX server behind a wrapper that points the agent at localhost instead of the cloud.
MoE, 256K native context, lowest hallucination of the three on factual-faithfulness leaderboards. Its censorship is political, not factual — it does not trigger on code/spec ingest.
Google (Western, no CCP political censorship); strongest agentic Tau2 (76.9%) of the dense Gemma 4 line; 256K native context.
Meta (Western); best open-weight hallucination score (~4.1%); 131K native context.
Verification stages
A model is only useful for the skills once it reaches Stage 2 — tool-calling — because the /memstead:* skills work by calling MCP tools. Inference alone (Stage 1) does not prove the skills run.
| Stage | Qwen3.6-35B | gemma-4-31b | Llama-3.3-70B |
|---|---|---|---|
| On disk, complete | ✓ | ✓ | ✓ |
| Discovered by oMLX | ✓ | ✓ | ✓ |
| Stage 1 — inference | ✓ | ✓ | ✓ |
| Stage 2 — tool-calling | ✓ in use | not run | not run |
Skills × models
Which Memstead skills have been run against which model. Tool-calling is the shared mechanism: once it's verified for a model, every skill rides the same path — but only an actually-run skill is marked verified.
| Skill | Qwen3.6-35B | gemma-4-31b | Llama-3.3-70B |
|---|---|---|---|
| /memstead:ingest | ◐ capable | candidate | candidate |
| /memstead:sync | ◐ capable | candidate | candidate |
| /memstead:verify | ◐ capable | candidate | candidate |
| /loop (unattended) | ◐ capable | candidate | candidate |
What's actually verified today: the one demonstrated tool-calling run is /memstead:ingest macos-graph on Qwen3.6 (May 2026) — it verifies Qwen's MCP tool-calling stage. That run predates the skill rewrite (the ingest router and the sync/verify skills were rebuilt since), so no individual skill carries a "verified" cell: every skill (ingest, sync, verify, /loop) rides the identical verified tool-calling path and is marked capable on Qwen — not claimed as separately benchmarked. Gemma and Llama run inference cleanly but haven't been put through tool-calling, so the skills are unverified on them: candidates, not portability claims.
This asset grows as runs land. It is deliberately not a launch gate — the launch ships whether this matrix is partial or complete.