Appearance
Unlocked native Messages qualification on November 9
The full repository gate and all six fresh desktop cases pass on the established Glimmer Q8 configuration at source commit 8ea41731. This closes the unlocked-gate and current-baseline desktop gaps. Experimental tiled execution remains off and unselected. The machine-readable record retains counts, timing boundaries, configuration, and evidence hashes.
Repository gate
The unmodified ./scripts/check.sh completes with exit code zero. It runs 5,919 XCTest cases, with 493 skips, and 197 Swift Testing cases. Every one of the 30 protected run-record cases that failed while locked passes while unlocked; file protection is unchanged. The official Anthropic SDK environment and installed Ornith/Glimmer tokenizer fixtures are enabled. Strict lint, build, CLI help, package boundaries, and hygiene checks pass.
These are CPU unit/contract checks. Optional checkpoint and GPU cases remain opt-in; the earlier numerical and native qualification evidence is retained.
Configuration and reference loop
Inference uses native mere.run Swift/MLX, the unchanged symmetric Q8/group-64 Glimmer target, and its BF16 DFlash2 assistant on the 128 GB M5 Max. Activations and KV remain BF16. The established selection is head4, three drafts, deferred scheduling, cost-checked fallback, and original normalization. No tiled mode, normalization fusion, or Q4 conversion is selected.
Requests retain the published sampled preset: temperature 1.0, top-p 0.95, top-k 64, and high effort. The context is 65,536 tokens and the advertised output limit is 4,096. Claude Code 2.1.316 owns the agent loop and tool execution. Cua Driver 0.38.0 uses standard permission mode and the unchanged 13-tool reference MCP configuration. Only the connection-helper executable path changes to the frozen production build; task prompts and tool behavior remain unchanged. No permission-policy or MCP-registration changes are made.
The browser trials use installed Chrome 157.0.8103.38, verified against the Google stable release notice. Each launches a separate temporary profile with the browser sandbox enabled. The model operates Chrome through Cua UI tools. Personal profiles, browser scripting, and direct method URLs are excluded from the task path. All model inference is local; the browser accesses the public MDN website. Nonessential Claude Code network traffic is disabled for the trials.
Independently verified outcomes
| Task | First fresh run | Second fresh run |
|---|---|---|
| Calculator 6 × 7 | Pass, 292.94 s | Pass, 245.04 s |
| Calculator 47 × 23 | Pass, 245.02 s | Pass, 261.94 s |
| MDN search → interface → getAll() | Pass, 389.37 s | Pass, 429.87 s |
Each Calculator trial starts a distinct process and has an observed initial zero. Independent final window state and screenshots verify both the expression and result: 6 × 7 = 42 or 47 × 23 = 1,081. The trusted verifier checks the Calculator executable and closes only the owned process after verification.
Each browser trial starts a distinct Chrome process and profile. Independent trace reconstruction verifies a visible Search click, the URLSearchParams query, the interface-result click, and the getAll() method-link click in that order. Final window state, page text, screenshot, and isolated browser history verify the method page and the answer about repeated and absent values. The trusted verifier closes only the owned browser after inspection. Model-declared success alone is insufficient.
The runtime records 90 completed native generation rounds and one failed generation. The client records one tool error. Both errors occur in the second browser case and are followed by successful recovery. One stream ends with an unparseable-tool-parameter error; the server withholds the invalid executable input, and Claude Code retries in buffered mode. The separate Cua focus error dispatches no key; subsequent UI actions complete the required task. The precise invalid parameter is not available in the validated-wire evidence.
These observations qualify the listed cases; they do not establish a general agent error rate or an error-free model.
Timing scope
The server reports 8.17 seconds loading and 41.04 seconds warmup before readiness; process launch to readiness is 49.52 seconds. Compilation and materialization remain in startup accounting. Those costs are excluded from client task wall times.
Client wall times include inference, reasoning, Cua actions, and client work. Native counter differences separate request load, prefill, and decode for completed requests. Decode includes every generated reasoning, text, and tool-input token in those completed requests. The failed generation has no phase/token accounting in these counters; its failure and retry remain included in client wall time. Observed native decode rates span 15.44–19.44 tokens/s. These unseeded agent trajectories and variable cache state are not a controlled speed comparison. The first Calculator run overlaps the end of CPU repository validation. Queue wait is not separately measured; requests are sequential, and session-boundary snapshots have zero active and queued requests and only the selected Q8 target loaded. Its draft companion belongs to that runtime.
Qualification boundaries
The core API evidence includes official SDK coverage of both API versions, HTTP authentication/error/cancellation tests, and the earlier signed reasoning, search, context-edit, cache, document, and image combinations. This follow-up changes no runtime source and does not promote experimental tiling.
Strict tools, constrained JSON-schema output, exact thinking budgets, generated citations, scanned-PDF rendering, and server compaction remain unsupported and rejected. Earlier extended-decimal and Gemma desktop-repeat failures remain in the qualification history. Success on these six cases does not qualify unlisted models, workflows, or hardware.
All six owned app processes, the native server, the task-owned Cua daemon, and the keep-awake helper are stopped after verification. Other worktrees and user processes remain untouched.