Homebrew installation: App quits unexpectedly on macOS Tahoe 26.4.1

[siathalysedI]

Description

Vesta crashes immediately on launch with a fatal error in MLXNN during model initialization. The application initializes the ContentView but terminates with a trace trap before any UI interaction is possible.

Steps to Reproduce

1. Install Vesta 0.9.7 via Homebrew: brew install --cask scouzi1966/afm/vesta-mac
2. Launch Vesta from Applications folder or via Terminal
3. Application initializes ContentView, then crashes during model loading phase
4. Error: Fatal error: 'try!' expression unexpectedly raised an error: MLXNN.UpdateError.unexpectedStructure

Expected Behavior

Application should launch successfully and allow model configuration/selection.

Actual Behavior

Application crashes on launch with MLXNN model structure error. Exit code: trace trap (SIGTRAP).

System Information

Field	Value
macOS Version	26.4.1 (Build 25E253) - macOS Tahoe
Architecture	arm64 (Apple Silicon, Mac14,6)
Vesta Version	0.9.7 (Build 10)
Installation Method	Homebrew Cask (scouzi1966/afm/vesta-mac)
Installation Date	2026-03-07
Apple Intelligence	Unavailable (language mismatch: system=en-GB, siri=en)

Crash Details

Terminal output when launching from CLI:

2026-05-12 23:47:20.404 Vesta[21408:23554538] VESTA_CONTENTVIEW_INIT: ContentView initializing (integrated mode with sidebar)
[1] 21408 trace trap  /Applications/Vesta.app/Contents/MacOS/Vesta 2>&1

Fatal error extracted from system logs:

libswiftCore.dylib) MLXNN/Module.swift:591: Fatal error: 'try!' expression unexpectedly raised an error: 
MLXNN.UpdateError.unexpectedStructure(key: "layers", item: "Model {
  embed_tokens: Embedding(embeddingCount=151936, dimensions=4096),
  layers: [
    DecoderLayer {
        input_layernorm: RMSNorm(dimensions=4096, eps=1e-06),
        mlp: MLP {
            down_proj: Linear(inputDimensions=12288, outputDimensions=4096, bias=false),
            gate_proj: Linear(inputDimensions=4096, outputDimensions=12288, bias=false),
            up_proj: Linear(inputDimensions=4096, outputDimensions=12288, bias=false),
          },
        post_attention_layernorm: RMSNorm(dimensions=4096, eps=1e-06),
        self_attn: Attention(headDim=128, heads=32, kvHeads=8, rotaryEmbedding=MLXVLM.Qwen3VLLanguage.RotaryEmbedding, scale=0.088388346)

Key indicators from crash report:

• Exception type: EXC_BREAKPOINT, signal: SIGTRAP
• Faulting thread: com.apple.root.user-initiated-qos.cooperative
• Error location: MLXNN/Module.swift:591 in swift_unexpectedError
• Model architecture detected: Qwen3-VL variant (embeddingCount=151936, dimensions=4096, MLP intermediate=12288)

Crash report file:

• Path: ~/Library/Logs/DiagnosticReports/Vesta-2026-05-12-234724.ips
• Incident ID: A244F5CF-006B-4167-ADAA-1BD08734FA89

Model Configuration

• Model attempted: Qwen3-VL variant (based on architecture signature)
• Backend: MLX (Apple Silicon native)
• Config file: ~/Library/Application Support/Vesta/config.json does not exist (first launch)

Troubleshooting Attempted

• Launched from Terminal to capture stdout/stderr
• Verified installation via brew info --cask vesta-mac
• Confirmed architecture match (arm64 binary on arm64 hardware)
• Reviewed crash report (.ips file)
• Cleared model cache (pending maintainer guidance)
• Reinstalled application (pending maintainer guidance)

Additional Context

• This is a fresh installation; no previous Vesta configuration existed on this system
• Other MLX-based applications work correctly on this system
• Error signature matches known issue pattern: MLXNN.UpdateError.unexpectedStructure during model deserialization
• Crash occurs before any user interaction, suggesting model validation happens at app launch rather than on first use
• Apple Intelligence is unavailable due to language mismatch (may be unrelated but noted for completeness)

Related Issues

• Potentially related to fix in v0.9.7 for "Jinja2 template bug causing crashes with Qwen3.6 and Gemma 4 models"
• If this is a model download/corruption issue, may need model re-download workflow or checksum validation

Request

1. Confirmation if this is a known issue with Qwen3-VL models in v0.9.7 on macOS Tahoe 26.4.1
2. Guidance on clearing corrupted model state to force re-download
3. If applicable, ETA for fix or workaround for MLXNN.UpdateError.unexpectedStructure
4. Clarification on which model files to remove if manual cleanup is required

Attachments Available

• Crash log: Vesta-2026-05-12-234724.ips (22KB)
• Terminal capture: vesta-error.log
• System info: sw_vers, uname -m, brew info --cask vesta-mac

[scouzi1966]

Thanks for the report. Which specific model did you attempt to load? If you can also include the hugging face reference of it as well.

[scouzi1966]

I am unable to reproduce this issue.