Midium Desktop
Changelog
What's new in each release. Midium updates from within the app; you can also run midium update.
Laguna S and Gemma 4 12B and Gemma 4 31B batching and speed improvements
Laguna S 2.1 now has efficient batching, along with improvements to Gemma 4 12B and Gemma 4 31B speculative decode and batching.
Models
- Laguna S 2.1 now is able to batch requests
- Laguna S 2.1 batches requests even more efficiently than stock
- Gemma 4 12B and 31B batched requests are much faster
- Gemma 4 12B and 31B speculative decoding is even faster
- Speculative decoding turns off conditionally and automatically when requests come in and out in a batched scenario
Other fixes
- Fixes for CLI to make it better coming from Courier to Midium
Qwen 3.8 Flash Next batching now available, along with new page features and redesigns!
This update includes an update to Qwen 3.8 Flash Next that makes it batching capable, along with stylistic and functionality improvements to several pages and models across Midium.
Models
- Qwen 3.8 Flash Next batching support
- Qwen 3.8 Flash Next adaptive MTP support in-conjunction with batching - on a deployed, batched model, MTP is enabled when one request is hitting the model, but automatically disengages when multiple requests hit the model at the same time
- Fixes to Mellum 2.1 sampling and thinking that enable it to perform much better
Pages
- Chat and Code pages now have different styling, along with interactive backgrounds that make the app more like Midium
- Studio page styling changed to match other pages container styling
- Vaults page has been refreshed with new vault cards that show your vault file distribution, along with a brief on what Midium Vaults do and how powerful they are
Qwen3.8 Flash Next on Midium is one of the fastest in the world on Mac!
This update provides even more throughput enhancements for Qwen3.8 Flash Next, as well as other models.
Models
- Qwen3.8 Flash Next is now among the fastest in the world on Midium among Macs
- Other model families such as other Qwen, Gemma 4, and Laguna S 2.1 models have also received a throughput upgrade
- Added support for Mellum 2.1 by Jetbrains
- Added support for LFM2.5 24B A2B by LiquidAI
Bug fixes
- A bug related to the memory override button is now fixed, with the ability
Qwen3.8 Flash Next even faster, with more throughput enhancements to come!
This update features even more throughput enhancements for Qwen3.8 Flash Next, including big improvements to prefill and uplifts to raw decode speed.
Models
- Throughput enhancements to Qwen3.8 Flash Next
- N-gram table streaming, resulting in Qwen3.8 Flash Next taking up around 75GB instead of 100GB+
Greatly improved Qwen3.8 Flash Next performance, boosted Gemma 4 family performances, boosted Laguna S 2.1 performance, and more!
This update features a ton of speed upgrades for all models.
Models
- Qwen3.8 Flash Next now is speedy quick at over 60 tok/s in normal conversation with boosts up to 240 tok/s in quick responses and ranges of speeds depending on teh work being done. We are currently working on updates that will make Qwen3.8 Flash Next the fastest in the world on Midium.
- Gemma 4 family, and especially Gemma 4 12B an Gemma 4 26B A4B have had updates to speed as well, also making these models among the fastest in the world on Midium.
- Laguna, specifically Laguna S 2.1, has had additional speed improvements; Laguna XS 2.1 untested, but probably faster.
- Conversational KV caching has been unlocked for Qwen, Gemma, and Laguna S 2.1 to allow much quicker responses; change your conversational KV cache settings in settings pertaining to your model on its model card on the Models page in Settings.
- Fixed a bug that caused Decider models to blow up memory size.
Record-breaking speed increases for Gemma 4 family, support for GLM 5.3 Flash, fixes for Qwen3.8 Flash Next, and more!
This update includes model updates, speed increases, and other big fixes that redefine how serious Midium is as positioning itself as a competitor in the model serving space on Mac.
Models
- Added support for GLM 5.3 Flash
- Increased throughput anywhere from 1.1x - 2x for Gemma 4 26B, Gemma 4 12B, and Gemma 4 31B
- Added conversational KV caching for Laguna S 2.1
- Increased throughput for Qwen3.8 Flash Next to be industry leading at over 50+ tok/s on a Macbook Pro 128gb M5 Max @ 4bit
- Added the ability to remove models from their 'Ready' state via a new Power Off button on model cards
CLI
- Fixed problem with the CLI not communicating with app shell when executing commands
- Added new
midium opencommand to open app when a current Midium process has been launched from CLI
Safety improvements
- Improved repetition detection to be more strict with repetitious tool calls and general LLM outputs while preventing false positives from being flagged
Models can now be imported from your file system
This update includes the ability to import models from your Midium models folder, along with other bug fixes found during testing.
Other feature additions and bug fixes
- Added the ability to modify the model file path
- Fixed a bug that could cause un-cancellable model downloads if quit during download
- Fixed a bug that could cause the "Remove model
UI/UX refactor across multiple pages, new tray popover, Qwen3.8 Flash Next and Bonsai 2 27B now supported, and much more!
The Midium UI/UX has been refactored in multiple places to make it much more intuitive from the get-go. Midium's tray icon has now been refactored with a custom popover. Qwen3.8 Flash Next and Bonsai 2 27B is now supported in Midium.
Models page
- Model cards and their settings have been refactored in grid and list views to be far more intuitive
- Added model logos to each model in the Catalog
API page
- Added reasoning effort support for supported models
- Added System One models available model types for testing purposes
- Added message scroller support and markdown formatter to response container for nicely formatted responses
Chat composer
- Consolidated multiple buttons and their functions in the Plus (+) menu into one
- Moved the shadow manager shield icon and the sub-agent icons next to the browser icon in the app header
- Added reasoning effort popover menu for models that are supported
Midium Hub invites
- Added share button that allows you to airdrop invite code to nearby macOS users
- Added copy button in invite field
Other features and bug fixes
- Added connection badge on the top right of the app that displays your current connection and tunnel status (if configured), along with quick access to other pages
- Enabled right clicking to allow macOS context menu to enable copy and paste within Midium
- Added deep linking and verification functionality for Always Awake so that it is easier to tell if things are set up correctly and how to diagnose issues
- Downloads now resume if the Midium is mid-download and is shut down
- Added shutdown overlay to display shutdown progress
Midium Hub demotions now possible
Midium Hub users can now demote themselves and release their license from their Mac all from within the app.
Onboarding fixes and Vault improvements
This update targeted bug fixes that users have been reporting during onboarding, as well as improvements to Midium's vault retrieval system so users with small Macs can still benefit from Midium's vault system.
Onboarding
- Fixed a bug that caused license registration to seemingly succeed, but then prompt a reactivation upon dashboard navigation.
- Fixed a bug that caused timeouts during license registration
Vaults
- Made Macs unable to run a reranker download and load Qwen3 0.6B instead BGE-M3 as our tests have shown 0.6B performs better reranker-less than BGE-M3.
- Capped Qwen embeddings models to a lower token limit to cap total memory usage.
General
- Updated light mode Midium icon to a darker color.
Initial vaults completed, and LAN functionality finished!
Midium now is at a state where teams can be connected through a Hub, with all of the supporting features that ensure collaboration! There are too many features to go over at the moment, since this is such a big update, but onboarding, Vault functionality, team functionality, general quality of life, and bug fixes are what's new.
Better RAG retrieval, application loading, and more
This update makes the capability of Midium to retrieve good context even better and more efficient, even with dense vaults.
Vaults
Midium features vaults, which role-aware and gated that allow you to ground a chat in to retrieve context pertinent to what you are working on. On the Free version, you are granted one vault that, when set up correctly, is automatically indexed and prepped for use with agents you use in Midium and sub-agents that other, external providers can use (Claude Code, Cursor, etc.). On a commercial plan, you gain access to have as many vaults as you want, with the ability to share those vaults as context with team members connected to your vault as Team Members. This update, retrieving the right context has gotten even better, featuring the ability to retrieve the right context 98 percent of the time in our in-house testing with over 500 documents featuring similar verbiage that really tests context awareness when executing RAG. We have added the ability to enter "company brain' mode, which grounds your chat in all vaults available to you.
Other feature additions and bug fixes
- Midium now launches smarter, waiting for the dashboard to fully compile and build before it starts rendering your dashboard.
- Fixed a bug with downloading that didn't accurately reflect the download process, along with better resuming of incomplete downloads.
- Consolidated all logs into one unified log that is accessible via the Logs page in Settings.
- Better performance in general.