Buckets:
| <rss xmlns:sparkle="http://www.andymatuschak.org/xml-namespaces/sparkle" version="2.0"> | |
| <channel> | |
| <title>Llama</title> | |
| <item> | |
| <title>0.47.0</title> | |
| <pubDate>Thu, 05 Nov 2026 13:25:34 +0200</pubDate> | |
| <sparkle:phasedRolloutInterval>7200</sparkle:phasedRolloutInterval> | |
| <sparkle:version>261105.1323</sparkle:version> | |
| <sparkle:shortVersionString>0.47.0</sparkle:shortVersionString> | |
| <sparkle:minimumSystemVersion>15.0</sparkle:minimumSystemVersion> | |
| <description><![CDATA[<html lang="en"><style>h2{font-size:inherit}ul{padding-left:1.5em;list-style:none}li::before{position:absolute;margin-left:-1.5em;content:"—"}time{color:hsl(from currentColor h s l / .25)}time::before{content:"· "}code{font-size:.85em}</style><h2>0.47.0 <time>Nov 5, 2026</time></h2><p>This release updates the engine to b12110, which adds support for d1 Omni, Liquid AI's tiny decision model. Give it text, images, audio or video and typed questions, and it returns a probability for each option.</p><ul><li>Update the engine to b12110, adding support for d1 Omni</li><li>Show when a model fails to load, and say when the engine may be too old</li><li>Use the app's own engine when your llama.cpp install is too old</li><li>Suggest a good first model at the top of Recommended for your Mac</li></ul><h2>0.46.1 <time>Oct 28, 2026</time></h2><p>On macOS 15, the engine in 0.46.0 couldn't load any model. This release fixes that.</p><ul><li>Fix models failing to load on macOS 15</li></ul><h2>0.46.0 <time>Oct 28, 2026</time></h2><p>This release updates the engine to b11941, which adds support for LFM3 Omni, Liquid AI's tiny model that understands text, images, and audio. It also runs Flint, Recursive's coding-agent model.</p><ul><li>Update the engine to b11941, adding support for LFM3 Omni and Flint</li><li>Open chat in your Safari Web App when you've added one to the Dock</li><li>Use the llama CLI from install.sh when it's at least the pinned version</li><li>Fix a second quant of a model missing image input or drafting</li><li>Keep a quant's image input when deleting another quant of the same model</li><li>Say a linked model is too large when none of its quants fit your Mac</li><li>Show the Meta logo for Muse models</li></ul><h2>0.45.0 <time>Oct 18, 2026</time></h2><p>Llama now runs embedding models, starting with EmbeddingGemma 2. They turn text into vectors for search and retrieval, through the server's <code>/v1/embeddings</code> endpoint. They carry an Embedding chip, and their model page links to a guide on how to use them.</p><ul><li>Sign in with Hugging Face to download gated models</li><li>Say when a gated model needs access instead of blaming the network</li><li>Add a Load row to the model page to warm a model up before use</li><li>Let decision models take inputs up to 4k tokens on every Mac</li><li>Update the engine to b11671, adding support for Nemotron 3.5 Super VL</li><li>Fit Qwen 3.8 Flash Next on 128 GB Macs</li><li>Show the Liquid AI logo for LFM models</li><li>Use standard-size controls and a one-line token field in Settings</li><li>Show focus on the shortcut recorder and cancel it on a click outside</li></ul><h2>0.44.0 <time>Oct 8, 2026</time></h2><p>Llama now runs decision models. Instead of chatting, they answer a typed question with a probability for each option, through the server's <code>/v1/systemone</code> endpoint. They carry a Decision chip, and their model page links to a guide on how to use them.</p><ul><li>Download DFlash draft heads when available: up to 50% faster than MTP</li><li>Update the engine to b11429, switching right away when no model is loaded</li><li>Merge the Downloads and Command settings tabs into a new Advanced tab</li><li>Rename the Web UI settings tab to Chat and its setting to custom web app</li><li>Open the quick prompt without bringing Settings to the front</li></ul><h2>0.43.0 <time>Oct 2, 2026</time></h2><p>You can now replace the built-in chat with your own web app. Pick a folder under Settings → Web UI and Llama serves it in place of the chat, ready to use your models with no extra setup. A different chat, a small tool for one task, an app for your phone: anything that runs in a browser.</p><ul><li>Show server state as a dot next to the menu title: green when it's serving</li><li>Show the QR code inline in the menu instead of in a separate popover</li><li>Show the base URL without /v1 so it works with more clients</li><li>Add a Server log row to Settings that opens the log in Console</li><li>Keep localhost working when the server is bound to Tailscale or an address</li><li>Update the engine to b11200, applies on next start</li></ul><h2>0.42.0 <time>Sep 2, 2026</time></h2><p>You can now let other devices reach the server. A new network access setting offers three options: <code>Off</code>, <code>Tailscale</code>, or <code>This network</code>. Tailscale is a free app that joins your devices into a private network, and it's the safest choice: your devices reach the server from anywhere, and nobody else on the local network can.</p><ul><li>Add a button that shows the server address as a QR code for phones to scan</li><li>Turn network access off when the Mac changes network or Tailscale signs out</li><li>Show the right server address on Macs whose main connection isn't Wi-Fi</li><li>Split Settings into General, Network, Downloads, Web UI and Command tabs</li><li>Show an MTP chip on installed models that draft ahead to generate faster</li><li>Update the engine to b10679, applies on next start</li></ul><h2>0.41.0 <time>Aug 18, 2026</time></h2><p>Every model now has a Build an API request row that opens a page for composing a call against your own server. Choose which API to speak — OpenAI chat, OpenAI Responses or Anthropic — turn thinking off, add a JSON schema, an image or a tool, and copy the result as curl, Python or JavaScript. Send it from the page to see the response.</p><p>It replaces Copy curl command, which handed you one fixed request and no way to change it.</p><p>You can also override the <code>models.ini</code> the app generates. Put your own settings in <code>~/.config/llama/models.user.ini</code> and they win, so you can pass arguments the app doesn't expose. A commented template is written there on first launch. If something in it can't be applied, nothing breaks and the menu tells you which overrides were ignored.</p><p>A download that hit a network hiccup could keep showing as downloading long after it finished, with the menu bar indicator stuck alongside it. It now clears.</p><h2>0.40.0 <time>Aug 11, 2026</time></h2><p>Quick prompt is a small panel you can summon from anywhere to type a prompt and send it straight to the web UI. It has no shortcut out of the box, so pick one in Settings under Web UI to start using it.</p><p>The default server port also changes from 8080 to 9931, following llama.cpp, which is preparing to switch its own default too. Scripts and API clients pointing at 8080 will stop working unless you update them, or set the port back to 8080 in Settings. A port you set yourself is untouched.</p><ul><li>Settings splits into General, Web UI and Advanced sections, with a sidebar</li><li>A "Source code" link now sits in the Settings sidebar</li><li>The server is ready right after an app update, instead of waiting on a download</li><li>Downloads show a live transfer rate, so large ones no longer look stuck</li><li>Opening the web UI from the menu loads the model immediately</li><li>Clicking a link no longer freezes the app while the browser starts</li><li>Failed downloads report as a warning instead of badging the app icon</li></ul><h2>0.39.0 <time>Aug 4, 2026</time></h2><p>Installing a vision model from Gemma 4, Qwen 3.6 or Devstral 2 quietly gave you a text-only model, with nothing to explain the missing image input. Those models now install with image support intact. Reinstall any already on your Mac that still can't see images.</p><p>Improved:</p><ul><li>Download progress fills the model chip instead of tracing a ring around it</li><li>A model's logo fades while its download is in flight</li><li>Laguna models show the Poolside logo</li><li>The llama engine moves to b10217</li></ul><p>Fixed:</p><ul><li>Vision models install with image support instead of loading text-only</li><li>Speech-to-text models no longer appear under Installed</li><li>A speculative draft head can no longer take the place of the real model</li><li>Models no longer appear twice in the menu</li><li>Leftover partial files are cleaned up once a model is installed</li></ul><h2>0.38.0 <time>Jul 27, 2026</time></h2><p>The memory budget is now measured from what your Mac reports rather than estimated from its total RAM, so bigger context sizes are available to pick. Models also offer every context length they were trained for, instead of stopping at 128k.</p><p>Other changes:</p><ul><li>A chevron marks the model rows that open a model's page when clicked</li><li>Pause, resume, and cancel a download from the row itself</li><li>A '?' by the context picker explains what a longer context costs you</li><li>Unload moved to the model page, and the Delete row drops its red tint</li></ul><h2>0.37.0 <time>Jul 24, 2026</time></h2><p>Click an installed model in the menu to open its own page -- disk size, a context-length picker with per-tier memory costs, and every per-model action in one place: chat, copy model ID, copy a curl command, view on Hugging Face, unload, and delete.</p><p>Other changes:</p><ul><li>Copy a ready-to-paste curl command for a model's API</li><li>View a model's Hugging Face repo from its page</li><li>A new setting to enable agent mode that passes --agent to the server</li><li>The context picker shows each tier's memory cost inline</li><li>Deleting a model now needs a confirming second click</li><li>The menu widens to fit long model names, up to a cap</li><li>A finished download no longer restarts the server or drops the loaded model</li></ul><h2>0.36.0 <time>Jul 21, 2026</time></h2><ul><li>Context picker lists every native tier, dimming ones your Mac can't run</li><li>Settings shows the app and engine versions and a GitHub link</li><li>Chat with a model is now one click from the row's hover buttons</li><li>Clearer copy when the one-time engine download fails</li><li>Suggestions skip families the catalog marks as low-memory picks</li></ul><h2>0.35.0 <time>Jul 14, 2026</time></h2><p>Menu rows now render model names as a parsed view of the model id -- a short name plus small chips for size and quant -- so a model looks consistent in the menu, the catalog, and the WebUI picker.</p><p>Breaking: API model ids now use <code>{org}/{repo}:{TAG}</code> for every org. Models from <code>ggml-org</code> no longer get shortened ids: <code>qwen3-0.6b:Q8_0</code> is now <code>ggml-org/Qwen3-0.6B-GGUF:Q8_0</code>. Update any API calls that use the old ids.</p><p>Other changes:</p><ul><li>Vision-capable models show a glasses marker in the menu and catalog</li><li>Models from non-default orgs show a dim <code>org/</code> prefix</li><li>Unload is now a hover action with an eject icon on the active model</li><li>Fix vision for repos with nonstandard <code>mmproj</code> filenames</li><li>Fix install links with <code>quant=</code> for prefixed and legacy quant labels</li><li>Update llama.cpp to <code>b9986</code></li></ul><h2>0.34.0 <time>Jul 6, 2026</time></h2><ul><li>Model names in the menu now match the ids used in the API</li><li>Recommended now shows every model size that fits your Mac</li><li>Browse models is now a standalone menu row</li><li>Download progress shows as a ring with pause/play inside</li><li>Restart link in the menu header when the server hits an error</li><li>MTP speedups now work for unsloth builds too</li><li>Fix chat links for models with <code>UD-</code> quants</li><li>Fix corrupted multi-part downloads for specific quants</li><li>Partial sharded downloads no longer show as installed</li><li>Update llama.cpp to <code>b9870</code></li></ul><h2>0.33.1 <time>Jul 3, 2026</time></h2><ul><li>fix installed models still showing in the recommended list</li><li>fix Gemma E2B and E4B models showing their raw repo name when installed</li></ul><h2>0.33.0 <time>Jul 2, 2026</time></h2><p>Models that ship a multi-token-prediction head, like Qwen3.6, now use it for speculative decoding — noticeably faster generation, with no setup. Works whether the head is embedded in the weights or comes as a sidecar file the app downloads alongside them.</p><p>New:</p><ul><li>Open the chat UI for a model straight from its expanded details</li><li>See the exact server command in Settings, with syntax highlighting</li><li>Menu bar icon badges while downloading and when a download finishes</li></ul><p>Improved:</p><ul><li>Downloads survive app restarts and resume as paused rows</li><li>Dropped connections pause the download and auto-resume when back online</li><li>Download rows show a progress bar and transfer speed</li><li>Long installed lists collapse behind a Show more toggle</li><li>Recommended models show their quant next to the name</li><li>Native ggml-org models drop the org prefix from their names</li><li>Settings redesigned: wider layout, clearer rows, restore-default buttons</li><li>Update llama.cpp to <code>b9835</code></li></ul><p>Fixed:</p><ul><li>Vision models scanned from the cache now load their projector file</li><li>A clear error when another app holds the server port, instead of a dead URL</li><li>Orphaned <code>llama serve</code> processes no longer block the next launch</li><li>Settings window centers correctly</li></ul><h2>0.32.0 <time>Jun 17, 2026</time></h2><p>LlamaBarn is now Llama. It's the same app with a new name, and your settings and downloaded models carry over automatically.</p><p>Because of the rename, this update isn't automatic -- download and install the new app.</p><p>Also in this release:</p><ul><li>Default server port is now <code>8080</code> to match llama.cpp, was <code>2276</code></li><li>Set a custom port in Settings, including <code>2276</code> if you relied on the old one</li><li>Deeplinks now use the <code>llama://</code> scheme; <code>llamabarn://</code> still works</li></ul><h2>0.31.1 <time>Jun 12, 2026</time></h2><p>The app now also accepts <code>llama://</code> links, ahead of the upcoming rename to Llama. Existing <code>llamabarn://</code> links continue to work.</p><h2>0.31.0 <time>Jun 11, 2026</time></h2><p>LlamaBarn used to bundle the llama.cpp engine inside the app. Now it shares one install with your command line: if you already have llama.cpp the app uses it, and if the app installs it you have it in your terminal too.</p><ul><li>Uses an existing llama.cpp from the install script at <a href="https://llama.app">https://llama.app</a> or Homebrew if you have one</li><li>Shows the in-use llama.cpp build in the footer</li><li>Ships as a <code>1.6 MB</code> .dmg (down from <code>7 MB</code>)</li></ul><p>We've also replaced the built-in catalog. Instead of a fixed list in the app, a new "Recommended for your Mac" section suggests models that fit, to get you started. For browsing, there's now a curated catalog at <a href="https://llama.app">https://llama.app</a> that's more up to date and easier to maintain.</p><p>Heads-up: a future update will rename LlamaBarn to Llama. It updates automatically and your models and settings carry over. Two things to know if you connect to the app: the local server will move to port <code>8080</code> (from <code>2276</code>) to match llama.cpp's default, and <code>llamabarn://</code> deep links will become <code>llama://</code>.</p><h2>0.30.0 <time>Apr 25, 2026</time></h2><ul><li>Add Qwen 3.6 family: 27B and 35B-A3B</li><li>Install models from Hugging Face via llamabarn:// deeplinks</li><li>Pause and resume in-progress downloads; partials survive app quit</li><li>Enable prompt-based speculative decoding by default</li><li>Find sideloaded models in HF cache subdirectories; fix split-shard quant labels</li><li>Fix MoE compatibility for sideloaded models using measured memory</li><li>Fix sideloaded estimation hanging forever when llama-fit-params failed</li><li>Improve sideload memory estimate accuracy</li><li>Move models.ini to Application Support; ~/.llamabarn no longer required</li><li>Update llama.cpp to b8902</li></ul><h2>0.29.1 <time>Apr 16, 2026</time></h2><p>Fix notarization. 0.29.0 shipped unnotarized due to a build pipeline bug.</p><h2>0.29.0 <time>Apr 15, 2026</time></h2><p>This release opens LlamaBarn up beyond the curated catalog: any GGUF model in your Hugging Face cache now shows up in the installed list with the same one-click load, run, and delete as curated models, with context tiers sized to your device automatically.</p><ul><li>Detect and support sideloaded GGUF models from the Hugging Face cache</li><li>Match llama-server format for sideloaded model IDs so IDs are portable</li><li>Default every model to the 4K context tier for a smaller footprint</li><li>Show the model's native max context alongside the device-fit tier</li><li>Show every size in catalog family drawers, with installed ones badged</li><li>Keep deprecated families like Qwen3 visible for already-installed models</li><li>Add a caption under Launch at login explaining idle resource use</li><li>Show friendlier HTTP download errors</li><li>Fix Gemma 4 download URLs after Hugging Face repo file renames</li><li>Update llama.cpp to b8797</li></ul><h2>0.28.0 <time>Apr 3, 2026</time></h2><ul><li>Add Gemma 4 models to catalog: 31B, 26B-A4B, E4B, E2B</li><li>Update llama.cpp to b8648</li></ul><h2>0.27.0 <time>Apr 1, 2026</time></h2><p>New downloads are now stored in ~/.cache/huggingface/hub/ using the standard Hugging Face cache layout. This means models downloaded by LlamaBarn can be used by llama.cpp and other Hugging Face aware tools. Existing models in ~/.llamabarn/ continue to work.</p><h2>0.26.0 <time>Mar 24, 2026</time></h2><ul><li>Add Qwen 3.5 models, replacing Qwen 3</li><li>Add Hugging Face token option to settings</li><li>Update llama.cpp to b8496</li></ul><h2>0.25.0 <time>Feb 18, 2026</time></h2><ul><li>Add custom models folder setting</li><li>Update llama.cpp to b8088</li></ul><h2>0.24.0 <time>Jan 28, 2026</time></h2><ul><li>Move settings to a separate window</li><li>Add context length variants for installed models</li><li>Add delete and copy model ID buttons on hover</li><li>Removed click-to-load as models now load on demand since 0.22</li><li>Add GLM 4.7 Flash to catalog</li><li>Update llama.cpp to b7843</li></ul><h2>0.23.0 <time>Jan 19, 2026</time></h2><ul><li>Allow binding server to a specific IP address</li><li>Set default sleep idle time to 5 minutes</li><li>Update llama.cpp to b7772</li></ul><h2>0.22.0 <time>Jan 15, 2026</time></h2><p>LlamaBarn now uses <code>llama-server</code> in Router Mode. The server stays running in the background and loads models automatically when they are needed. You no longer have to manually select a model before using it. This version also adds an optional "Unload when idle" setting that automatically removes models from memory after a period of inactivity.</p><ul><li>Migrate to llama-server Router Mode</li><li>Introduce "Unload when idle" setting for memory management</li><li>Add automatic download retry logic with exponential backoff</li><li>Add error reporting for failed model downloads</li></ul><h2>0.21.0 <time>Jan 7, 2026</time></h2><ul><li>Update header to show base URL as a separate element</li><li>Include memory budget details in context length info in Settings</li><li>Add Hugging Face link to model context menu</li><li>Remove --no-mmap to enable memory mapping</li><li>Update llama.cpp to b7652</li><li>Fix Nemotron KV cache footprint calculation</li></ul><h2>0.20.0 <time>Jan 2, 2026</time></h2><ul><li>Refine catalog menu layout and navigation</li><li>Show granular download progress with decimal percentages</li><li>Optimize settings menu toggle to avoid full menu rebuild</li></ul><h2>0.19.0 <time>Dec 30, 2025</time></h2><ul><li>Family items now open detailed views instead of expanding in place</li><li>Add descriptions to model families</li><li>Show incompatible models in the catalog with clear memory requirements</li><li>Group Qwen3, Qwen3 VL, and Ministral 3 models with their reasoning variants</li></ul><h2>0.18.0 <time>Dec 29, 2025</time></h2><ul><li>Add Q4_K_M quantizations for Devstral 2 models</li><li>Refactor catalog and model structures for better organization</li><li>Update llama.cpp to b7569</li></ul><h2>0.17.0 <time>Dec 28, 2025</time></h2><ul><li>Add Devstral 2 model family</li><li>Add Nemotron Nano 3 model family</li><li>Use xcassets for icon to reduce bundle size</li><li>Update llama.cpp to b7561</li></ul><h2>0.16.0 <time>Dec 19, 2025</time></h2><ul><li>Add Show in Finder button to installed model context menu</li><li>Always show context length and estimated memory usage in model items</li><li>Remove memory limit option in favor of automatic safety budget</li><li>Add expandable info descriptions to settings</li><li>Fix UI flicker when downloading or cancelling models</li><li>Update llama.cpp to b7475</li></ul><h2>0.15.0 <time>Dec 17, 2025</time></h2><ul><li>Added "Memory cap" setting to replace hard-coded logic and allow user control</li><li>Removed confusing right-click button for "run at max ctx"</li><li>Renamed "Context length" option to "Context length cap" and updated labels</li><li>Updated context length display in model metadata</li><li>Made model deletion immediate for better UX</li></ul><h2>0.14.0 <time>Dec 16, 2025</time></h2><ul><li>Removed <code>show quantized</code> toggle — catalog now prefers full-precision models and falls back to quantized only when full-precision won't run on device</li><li>Added setting to set context length</li><li>Added setting to display estimated memory usage</li><li>Redesigned model items and settings</li><li>Updated icon for better integration with macOS Tahoe</li><li>Updated llama.cpp to b7406</li></ul><h2>0.13.0 <time>Dec 4, 2025</time></h2><ul><li>Added Ministral 3 model family</li><li>Added vision support indicator to catalog</li><li>Replaced hover-based delete with right-click delete for installed models</li><li>Added toggle button for quantized models in catalog divider</li><li>Changed default context window from max to 4k</li><li>Improved visual hierarchy with refined colors and font sizes throughout</li><li>Updated app icon to better match menu bar icon</li><li>Updated llama.cpp to b7247</li></ul><h2>0.12.0 <time>Nov 25, 2025</time></h2><ul><li>Enable resuming model downloads after network interruptions</li><li>Add default sampling parameters based on model author recommendations</li><li>Add experimental option to expose llama-server to network by binding to 0.0.0.0 -- see issue #17</li></ul><h2>0.11.0 <time>Nov 13, 2025</time></h2><ul><li>Show welcome popover on first launch to guide new users</li><li>Fix app hanging issues by moving file I/O to background queues</li><li>Remove redundant update check at app launch</li><li>Make version number clickable to check for updates</li></ul><h2>0.10.0 <time>Nov 5, 2025</time></h2><ul><li>Add Qwen3-VL and Qwen3-VL Thinking vision-language model families with automatic mmproj detection</li><li>Standardize model naming in catalog with activated parameter counts for MoE models (e.g., 30B-A3B)</li><li>Improve code organization and catalog structure</li><li>Update llama.cpp to b6942</li></ul><h2>0.9.1 <time>Oct 29, 2025</time></h2><ul><li>Fix dock icon briefly appearing on app launch</li><li>Fix server URL not showing when model name is long</li></ul><h2>0.9.0 <time>Oct 23, 2025</time></h2><ul><li>Update llama.cpp to b6827</li><li>Update Sparkle to 2.8.0</li><li>Add Sentry error reporting for release builds</li><li>Simplify codebase by extracting common patterns</li><li>Remove unused code</li></ul><h2>0.8.1 <time>Oct 17, 2025</time></h2><ul><li>Fixed quantized model sizes not showing in collapsed family headers</li></ul><h2>0.8.0 <time>Oct 17, 2025</time></h2><ul><li>Reduce menu width from 320 to 300 points</li><li>Remove DeepSeek from catalog</li><li>Update appcast URL to releases subdomain</li></ul><h2>0.7.0 <time>Oct 15, 2025</time></h2><ul><li>Preserve family header highlight during collapse/expand interactions</li><li>Preserve family collapse state when toggling settings</li><li>Fix menu width changing when expanding/collapsing families</li><li>Hide trailing indicators in catalog models and expanded family headers to reduce visual clutter</li></ul><h2>0.6.0 <time>Oct 14, 2025</time></h2><ul><li>Add collapsible family headers in catalog (click to expand/collapse)</li><li>Remove catalog-level toggle in favor of family-level collapsing</li><li>Show model sizes inline when catalog family headers are collapsed</li><li>Start catalog families collapsed by default</li><li>Revert icon backgrounds to default color</li></ul><h2>0.5.0 <time>Oct 14, 2025</time></h2><ul><li>Add collapsible catalog section with persistent state</li><li>Add family headers to catalog section for better organization</li><li>Add model family colors for icon backgrounds</li><li>Update llama.cpp to b6756</li><li>Improve model size display formatting</li></ul><h2>0.4.0 <time>Oct 14, 2025</time></h2><ul><li>Simplified catalog UI by replacing submenus with a flat list of models</li><li>Hide incompatible models from catalog entirely instead of showing them grayed out</li><li>Show running model name in header with green highlight</li><li>Hide Installed section when empty (like Available section)</li><li>Improved icon and typography styling with refined small caps and alignment</li><li>Fixed catalog section visibility when all models are installed/deleted</li></ul><h2>0.3.0 <time>Oct 10, 2025</time></h2><ul><li>Improved menu hover performance by optimizing highlight tracking</li><li>Changed context window display to use binary units (128k instead of 131k for 131,072 tokens)</li><li>Updated context window capping display from "12k of 128k" to "12k capped"</li><li>Changed to rounded square icons for installed models</li><li>Replaced bullet separator (•) with middle dot (·) in metadata</li><li>Updated status icons in family items to match catalog items</li><li>Changed size formatting to omit decimal point when zero (4 GB instead of 4.0 GB)</li></ul><h2>0.2.0 <time>Oct 9, 2025</time></h2><ul><li>Updated llama.cpp to b6713</li><li>Installed models now grouped by family for easier comparison</li><li>Context window display improved: shows "32k of 128k" format when reduced due to memory constraints instead of strikethrough</li><li>Delete button now always visible on installed models instead of requiring hover</li><li>Fixed text wrapping in model family descriptions</li></ul><h2>0.1.0 <time>Oct 7, 2025</time></h2><p>Hello world 🦙 🌾</p>]]></description> | |
| <enclosure url="https://github.com/ggml-org/Llama-macOS/releases/download/0.47.0/Llama.dmg" length="1292425" type="application/octet-stream" sparkle:edSignature="BAOgnuFaZVJ5b8WPvO75FujCti0JDvXyGyiMXFq2QHucrjFCEOs040j1Acsv4gCBP/d0KlQ7uY/l5G8RiKbtDQ=="/> | |
| </item> | |
| </channel> | |
| </rss> |
Xet Storage Details
- Size:
- 27.3 kB
- Xet hash:
- 8c1e5591b0b4d7dc55d84484d95133bb1711ac2449e530fd1e1a9b68e56f93d4
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.