PocketModel keeps model choice and private local chat primary. Discover uses a readable adaptive layout across iPhone, iPad, and Mac, while catalog filters, diagnostics, and maintenance stay disclosed until needed. Model provenance, device fit, runtime fallback, and local privacy remain visible.

PocketModel keeps model choice and private local chat primary. Discover uses a readable adaptive layout across iPhone, iPad, and Mac, while catalog filters, diagnostics, and maintenance stay disclosed until needed. Model provenance, device fit, runtime fallback, and local privacy remain visible.

Version 2.6.11 build 9 simplifies the default local-chat flow. Discover now leads with Apple Foundation Models where available or a recommended local GGUF model, keeps catalog filters behind one Refine control in Simple mode, and moves storage/catalog maintenance details out of the way unless they need attention. Library now focuses on models that need action and models ready to use before maintenance details.

Version 2.6.10 build 8 refreshes the curated local model catalog with Gemma 4 E2B QAT, Ministral 3 Reasoning 3B, and the official IBM Granite 4.0 H Tiny GGUF artifact. It also adds device-fit filtering, clearer runtime budget and fallback status, persisted load evidence, and a redacted diagnostics export with expanded local privacy copy.

Version 2.6.8 build 4 completes the post-baseline GGUF catalog and review-readiness lane: - Adds a release-completion audit for runtime fallback proof, catalog checksum remediation, signing and rollback surfaces, document/RAG workflows, performance controls, privacy tools, accessibility, search, memory, and local observability - Keeps PocketModel-managed voice input and in-app camera capture hidden while permission rewrites and real-device validation remain gated - Preserves catalog trust, local runtime controls, screenshot/reviewer-note posture, and macOS signing blocker evidence in release docs - Keeps uncensored models behind Complete mode and an explicit Settings unlock - Preserves the existing verified catalog and local-only defaults

Version 2.6.8 build 3 completes the local-safe P0/P1 release-readiness lane: - Adds a release-completion audit for runtime fallback proof, catalog checksum remediation, signing and rollback surfaces, document/RAG workflows, performance controls, privacy tools, accessibility, search, memory, and local observability - Keeps PocketModel-managed voice input and in-app camera capture hidden while permission rewrites and real-device validation remain gated - Preserves catalog trust, local runtime controls, screenshot/reviewer-note posture, and macOS signing blocker evidence in release docs - Keeps uncensored models behind Complete mode and an explicit Settings unlock - Preserves the existing verified catalog and local-only defaults

Version 2.6.7 build 2 ships a PocketModel stability and release-validation update: - Supersedes the submitted iOS 2.6.4 build with the PocketModel deep-dive P0/P1 stabilization tranche - Keeps Camera capture and PocketModel-managed voice input hidden while hardware permission paths remain under validation - Improves compact chat action buttons and text-to-speech speak/stop reliability - Prevents hidden thinking output from appearing in titles, streamed text, search previews, exports, and spoken output when thinking is off - Improves iPad chat workspace navigation and adds native macOS commands for common actions - Strengthens download remediation, catalog trust policy, and release validation documentation - Adds fresh iPad M4 and iPhone 16 hardware-release evidence to the review train - Keeps uncensored models behind Complete mode and an explicit Settings unlock - Preserves the existing verified catalog and local-only defaults

Version 2.6.6 build 1 ships a PocketModel stability and platform polish release: - Supersedes the submitted iOS 2.6.4 build with the PocketModel deep-dive P0/P1 stabilization tranche - Keeps Camera capture and PocketModel-managed voice input hidden while hardware permission paths remain under validation - Improves compact chat action buttons and text-to-speech speak/stop reliability - Prevents hidden thinking output from appearing in titles, streamed text, search previews, exports, and spoken output when thinking is off - Improves iPad chat workspace navigation and adds native macOS commands for common actions - Strengthens download remediation, catalog trust policy, and release validation documentation - Keeps uncensored models behind Complete mode and an explicit Settings unlock - Preserves the existing verified catalog and local-only defaults

Version 2.6.5 build 1 ships a PocketModel stability and platform polish release: - Supersedes the submitted iOS 2.6.4 build with the PocketModel deep-dive P0/P1 stabilization tranche - Keeps Camera capture and PocketModel-managed voice input hidden while hardware permission paths remain under validation - Improves compact chat action buttons and text-to-speech speak/stop reliability - Prevents hidden thinking output from appearing in titles, streamed text, search previews, exports, and spoken output when thinking is off - Improves iPad chat workspace navigation and adds native macOS commands for common actions - Strengthens download remediation, catalog trust policy, and release validation documentation - Keeps uncensored models behind Complete mode and an explicit Settings unlock - Preserves the existing verified catalog and local-only defaults

Version 2.6.4 build 1 ships a focused PocketModel stabilization release: - Supersedes 2.6.2 with safer camera and voice permission handling - Keeps camera capture hidden until physical-device permission testing is available, while Files and Photos image import remain available - Refines chat action buttons with a lighter compact visual style and reliable text-to-speech toggling - Prevents hidden thinking output from leaking into automatic chat titles and disabled-thinking search previews - Keeps full-history regression audit evidence, measured performance proxies, and strict physical-device perf evidence handling - Keeps uncensored models behind Complete mode and an explicit Settings unlock - Keeps PocketModel-managed voice input disabled with a build-level safety switch while the permission path is rebuilt - Preserves the existing verified catalog and local-only defaults

Version 2.6.3 build 1 ships a focused PocketModel stabilization release: - Supersedes 2.6.2 with safer camera and voice permission handling - Keeps camera capture hidden until physical-device permission testing is available, while Files and Photos image import remain available - Improves chat action button hit targets and text-to-speech toggling - Prevents hidden thinking output from leaking into automatic chat titles and disabled-thinking search previews - Keeps full-history regression audit evidence, measured performance proxies, and strict physical-device perf evidence handling - Keeps uncensored models behind Complete mode and an explicit Settings unlock - Keeps PocketModel-managed voice input disabled with a build-level safety switch while the permission path is rebuilt - Preserves the existing verified catalog and local-only defaults

Version 2.6.2 build 1 ships a focused PocketModel stabilization release: - Supersedes 2.6.1 with a chat cancellation fix so stopped replies finalize cleanly and follow-up prompts remain available - Adds full-history regression audit evidence for crash-prone and action-prone areas - Adds measured release performance proxies for catalog/persistence warmup and formatter throughput - Hardens the physical-device perf guard so missing XCTest performance metrics no longer pass silently - Focuses the release on responsiveness, crash prevention, and incorrect-action regression coverage - Keeps uncensored models behind Complete mode and an explicit Settings unlock - Keeps PocketModel-managed voice input disabled with a build-level safety switch while the permission path is rebuilt - Preserves the existing verified catalog and local-only defaults

Version 2.6.1 build 1 ships a focused PocketModel stabilization release: - Adds full-history regression audit evidence for crash-prone and action-prone areas - Adds measured release performance proxies for catalog/persistence warmup and formatter throughput - Hardens the physical-device perf guard so missing XCTest performance metrics no longer pass silently - Focuses the release on responsiveness, crash prevention, and incorrect-action regression coverage - Keeps uncensored models behind Complete mode and an explicit Settings unlock - Keeps PocketModel-managed voice input disabled with a build-level safety switch while the permission path is rebuilt - Preserves the existing verified catalog and local-only defaults

Version 2.6.0 build 1 ships a broader local-first PocketModel stabilization release: - Adds Phi-4 Mini Reasoning for math, code, and step-by-step reasoning - Adds Gemma 4 E4B as a premium text model for large-memory devices - Adds local diagnostics, audit, retention, recommendation, and handoff foundations - Improves accessibility, privacy controls, model recommendation guidance, and release evidence - Keeps uncensored models behind Complete mode and an explicit Settings unlock - Keeps PocketModel-managed voice input disabled with a build-level safety switch while the permission path is rebuilt - Hardens user-added model IDs, remote catalog validation, full catalog verification, and catalog trust handling

Version 2.5.1 build 2 focuses on App Review stability and a simpler, safer local model catalog: - Fixes a crash when tapping the microphone button in Chat on iPhone, iPad, and Mac - Improves the Chat + action menu on iPhone and iPad so actions remain available during review - Simplifies the curated catalog around current high-quality Qwen-first models, with one advanced uncensored option kept behind Complete mode - Makes device-based model recommendations more conservative to reduce load and generation crashes - Hardens model switching, thinking-mode transitions, checksum diagnostics, and internal tag cleanup in assistant replies

Version 2.5.0 hardens the local model catalog: - Corrects downloadable-file checksums for affected Gemma, Llama, and Qwen Coder catalog entries - Refreshes displayed download sizes for recently updated catalog models - Adds release-gated Hugging Face metadata validation across the full bundled catalog - Keeps checksum diagnostics available when a local artifact is unverified - Continues running chat generation on-device with no account required

Version 2.1.1 adds document context to chat: - Attach PDFs, text files, or Markdown notes to any chat — the model sees the content as context - Manage your document library from the toolbar; import from Files with a tap - Source citations appear in the reply bubble so you can see which document informed the answer - Discover now organizes models by family (Qwen, Llama, Mistral, etc.) for easier comparison - Advanced: experienced users can long-press any model in Discover to browse and install GGUF files directly from its HuggingFace repository

Version 2.1.0 adds document context to chat: - Attach PDFs, text files, or Markdown notes to any chat — the model sees the content as context - Manage your document library from the toolbar; import from Files with a tap - Source citations appear in the reply bubble so you can see which document informed the answer - Discover now organizes models by family (Qwen, Llama, Mistral, Gemma, etc.) for easier comparison - Advanced: experienced users can long-press any model in Discover to browse and install GGUF files directly from its HuggingFace repository

Version 2.0.x improves the default on-device chat experience: - Added prompt suggestions for new chats and helpful follow-up suggestions after replies - Added chat search across the current thread and across saved conversations - Added edit, regenerate, and fork-new-chat flows for earlier prompts - Improved reply rendering so lists, links, and code formatting stay easier to read - Continued polishing onboarding, model recommendations, and privacy-first messaging

Version 2.0.2 improves the default on-device chat experience: - Added prompt suggestions for new chats and helpful follow-up suggestions after replies - Added chat search across the current thread and across saved conversations - Added edit, regenerate, and fork-new-chat flows for earlier prompts - Improved reply rendering so lists, links, and code formatting stay easier to read - Continued polishing onboarding, model recommendations, and privacy-first messaging

Version 2.0.2 improves the default on-device chat experience: - Added prompt suggestions for new chats and helpful follow-up suggestions after replies - Added chat search across the current thread and across saved conversations - Added edit, regenerate, and fork-new-chat flows for earlier prompts - Improved reply rendering so lists, links, and code formatting stay easier to read - Continued polishing onboarding, model recommendations, and privacy-first messaging

Version 2.0.1 focuses on stability and polish for the hybrid on-device runtime and adds Apple Foundation Models support and expands the runtime to a hybrid default: - Fixed a regression where Thinking Mode could remain enabled after switching to a model that does not support it - Improved onboarding and model-activation reliability for local GGUF installs - Refined chat transcript formatting and streaming stability during longer replies - Apple Foundation Models — on supported Apple Intelligence devices, PocketModel now uses the built-in system model by default, with no download required - Onboarding defaults to Foundation Models where available; Discover shows Foundation Models in a built-in section alongside the downloadable catalog - Settings Runtime section now reflects which provider is active (built-in Foundation Models or a specific loaded GGUF model) - Improved download reliability: model downloads now automatically retry via a mirror when the primary source has a transport failure

Version 2.0.0 adds Apple Foundation Models support and expands the runtime to a hybrid default: - Apple Foundation Models — on supported Apple Intelligence devices, PocketModel now uses the built-in system model by default, with no download required - Curated GGUF models remain fully supported as the fallback on all devices and the upgrade path for users who want explicit model choice - Onboarding defaults to Foundation Models where available; Discover shows Foundation Models in a built-in section alongside the downloadable catalog - Settings Runtime section now reflects which provider is active (built-in Foundation Models or a specific loaded GGUF model) - Improved download reliability: model downloads now automatically retry via a mirror when the primary source has a transport failure - All prior features from v1.2 — device-aware recommendations, Experienced mode, Thinking Mode, Spotlight, model update badges, download consent — continue to work as expected

Version 1.2.0 makes model selection clearer and daily chat use faster: - Discover now shows the full catalog, with a separate "Not Recommended for This Device" section for more demanding models - Added Standard and Experienced modes so advanced users can intentionally override device-fit recommendations after a warning - Model actions are clearer, including Download, Download Anyway, and Unavailable states with device-fit guidance - Onboarding now keeps recommended models front and center while letting experienced users reveal more demanding options - Inference presets now live in Settings, while Chat keeps starter actions and a composer action menu for Try Another Model and future secondary actions - Removed the app-managed voice-input path in favor of the iOS keyboard's built-in dictation for a more stable chat experience - Hardened model switching so active chats wait for the new model to finish loading before send or rerun can continue - Fixed a regression where a newly generated assistant reply could lose its model label after the first chat update - Fixed a generation-replacement race so canceling a previous response no longer interrupts the replacement response - Improved completed-response formatting so finished replies preserve the readable structure users saw while streaming, copying, and exporting - Added top/bottom jump controls for long chats, contextual follow-up reply chips, and tap-to-dismiss keyboard behavior in chat - Corrected the Settings version display so it matches the app bundle metadata

Version 1.1.0 brings a major model expansion, smarter device-aware recommendations, and quality-of-life improvements: - 13 verified models from 6 families — including Qwen 3.5 (with thinking mode), Ministral, Phi-4 Mini, SmolLM2, Nemotron Mini, and Mimo - Thinking Mode — Qwen 3.5 models now show chain-of-thought reasoning in a collapsible section, with a toggle in Settings - Device-aware model filtering — only models compatible with your device are shown during setup; incompatible models are clearly marked in Discover - Improved text formatting — chat responses now preserve line breaks and render Markdown correctly - Model label on responses — each assistant message shows which model generated it - Split-download reliability fix — large models that download in multiple parts now verify each part individually - Spotlight, text-to-speech, model update badges, and download consent from v1.0 continue to work as expected