Releases and downloads.

What's new

Chat: paste or drop images into a message (needs the model's vision projector — add --mmproj <file> to its flags)

Image and video generation is now a mode of Chat's mode menu, with its options in the composer and an image-model picker; the separate tab is gone

New Quantize card on a model: make a smaller GGUF from a BF16/F16 model with llama-quantize, with progress and cancel

Basic settings above the raw flags: context, KV cache quality, flash attention, batch, threads and the KVMem budget, with a live VRAM estimate

Auto-tune measures properly: warm-up, three runs on a ~1.5k-token prompt (median), free-VRAM check, KVMem-aware variants that build on each other

KVMem updates: the newest prebuilt release (including builds like rc3-prism.3) is now detected correctly

Remote chat from a phone: model picker, optional agent mode and an optional internet tunnel

Fixed: saving flags for a stopped server used to start it, and a second Start loaded two copies at once; Quick Start and Stop now follow the engine chosen for the model

Settings restyled with a dark look and the title in the top bar

Co nowego

Czat: wklejaj lub przeciągaj obrazy do wiadomości (wymaga projektora wizji modelu — dopisz do jego flag --mmproj <plik>)

Generowanie obrazów i wideo jest teraz trybem w menu trybów czatu, z opcjami w polu wiadomości i wyborem modelu obrazkowego; osobna zakładka zniknęła

Nowa karta Kwantyzacja przy modelu: zrób mniejszy plik GGUF z modelu BF16/F16 przez llama-quantize, z postępem i anulowaniem

Ustawienia podstawowe nad surowymi flagami: kontekst, jakość cache KV, flash attention, batch, wątki i budżet KVMem, z żywym szacunkiem VRAM

Auto-tune mierzy rzetelnie: rozgrzewka, trzy przebiegi na prompcie ~1,5 tys. tokenów (mediana), kontrola wolnego VRAM, warianty uwzględniające KVMem, które się na siebie nakładają

Aktualizacje KVMem: najnowsze gotowe wydanie (także takie jak rc3-prism.3) jest teraz poprawnie wykrywane

Czat zdalny z telefonu: wybór modelu, opcjonalny tryb agenta i opcjonalny tunel internetowy

Naprawiono: zapis flag zatrzymanego serwera uruchamiał go, a drugi Start ładował dwie kopie naraz; Quick Start i Stop zależą teraz od silnika wybranego dla modelu

Ustawienia w ciemnym stylu, z tytułem na górnym pasku

What's new

macOS and Linux installers are available again — 0.1.65's release build failed for those two platforms, so only Windows got it

Co nowego

Instalatory macOS i Linux znów dostępne — build wydania 0.1.65 padał dla tych dwóch platform, więc trafił tylko Windows

Llama Control v0.1.66

What's new

Linux (AppImage) is now a build target alongside Windows and macOS

Image generation: install stable-diffusion.cpp and generate images from local models, right in the app

Video generation (Wan2.1 T2V) on the same stable-diffusion.cpp engine as image generation

KVMem (experimental) — a second inference engine alongside llama.cpp for NVIDIA/CUDA GPUs, installed to its own versioned folder, never mixed with llama.cpp files

Voice mode: push-to-talk in Chat via whisper.cpp (speech-to-text) and llama-tts (text-to-speech), installable from Settings

Agent mode switch in Chat: flip a message between plain chat and an OpenCode agent turn with tool/file access, without leaving the conversation

New Chat diagnostics panel: filterable log of what a request actually did (errors, tool calls, permissions)

HuggingFace tab rewritten as a dark two-pane layout with a model README preview

New Poradnik (Guide) tab for connecting OpenCode/Crush

Sidebar redesigned: Models folded into a consolidated Settings tab, and collapsing it now reclaims the full width instead of leaving an icon rail

Sidebar's collapse button moved out of the window's top-left corner, so it no longer sits where macOS puts its own window controls

Co nowego

Linux (AppImage) jest teraz celem builda obok Windows i macOS

Generowanie obrazów: zainstaluj stable-diffusion.cpp i generuj obrazy z lokalnych modeli bezpośrednio w aplikacji

Generowanie wideo (Wan2.1 T2V) na tym samym silniku stable-diffusion.cpp co generowanie obrazów

KVMem (eksperymentalny) — drugi silnik wnioskowania obok llama.cpp dla kart NVIDIA/CUDA, instalowany do osobnego, wersjonowanego folderu, nigdy mieszany z plikami llama.cpp

Tryb głosowy: push-to-talk w Czacie przez whisper.cpp (mowa na tekst) i llama-tts (tekst na mowę), instalowalny z Ustawień

Przełącznik trybu agenta w Czacie: zamień wiadomość między zwykłym czatem a turą agenta OpenCode z dostępem do narzędzi/plików, bez wychodzenia z rozmowy

Nowy panel diagnostyki czatu: filtrowany log tego, co faktycznie zrobiło zapytanie (błędy, wywołania narzędzi, uprawnienia)

Zakładka HuggingFace przepisana na ciemny układ dwupanelowy z podglądem README modelu

Nowa zakładka Poradnik do podłączenia OpenCode/Crush

Przeprojektowany panel boczny: Modele przeniesione do skonsolidowanej zakładki Ustawienia, a zwinięcie panelu zwalnia teraz całą jego szerokość zamiast zostawiać pasek z ikonami

Przycisk zwijania panelu bocznego przeniesiony z lewego górnego rogu okna, żeby nie kolidował z natywnymi przyciskami okna na macOS

What's new

Fixed llama.cpp install and update: llama.cpp changed how it publishes releases, so the app could no longer find a build or detect the installed version

Updating llama.cpp now stops a running server first, so the files can be replaced

Co nowego

Naprawiono instalację i aktualizację llama.cpp: llama.cpp zmieniło sposób publikowania wydań, przez co aplikacja nie mogła znaleźć builda ani rozpoznać zainstalowanej wersji

Aktualizacja llama.cpp zatrzymuje teraz działający serwer, aby można było podmienić pliki

What's new

macOS (Apple Silicon) support — Llama Control now runs natively on M1 and newer Macs

Chat with your models from your phone: turn on "Remote access" in Settings and pair over your home Wi-Fi with a QR code or PIN, nothing to install on the phone

New Plugins tab, starting with a JSON Extractor — paste text, define a schema (or have the model suggest one from the text), and get back structured JSON guaranteed to match it

A "Connect a coding tool" button in Chat generates a ready-to-paste OpenCode/Crush config pointing at your local server

A setup screen now appears on launch if llama.cpp isn't installed yet, instead of leaving the app looking broken

The sidebar can now be collapsed to a slim icon rail to reclaim screen space

HuggingFace search now loads more results as you scroll, and shows each model's added date

Added a refresh button to the saved chats list

Noted why mouse scroll can stop working inside the OpenCode terminal panel, with the config fix

Co nowego

Wsparcie dla macOS (Apple Silicon) — Llama Control działa teraz natywnie na Makach z M1 i nowszych

Czat z modelami z telefonu: włącz „Zdalny dostęp" w Ustawieniach i sparuj urządzenia w tej samej sieci Wi-Fi kodem QR lub PIN-em — na telefonie nie trzeba niczego instalować

Nowa zakładka Wtyczki, a w niej Ekstraktor JSON — wklej tekst, zdefiniuj schemat (lub pozwól modelowi zaproponować go na podstawie tekstu) i otrzymaj ustrukturyzowany JSON gwarantowanie zgodny ze schematem

Przycisk „Podłącz narzędzie kodujące" w Czacie generuje gotową do wklejenia konfigurację OpenCode/Crush wskazującą na Twój lokalny serwer

Przy starcie pojawia się teraz ekran konfiguracji, jeśli llama.cpp nie jest jeszcze zainstalowane, zamiast pozostawiać aplikację w niedziałającym stanie

Panel boczny można teraz zwinąć do wąskiego paska z ikonami, by zyskać więcej miejsca na ekranie

Wyszukiwanie HuggingFace doładowuje teraz kolejne wyniki podczas przewijania i pokazuje datę dodania każdego modelu

Dodano przycisk odświeżania na liście zapisanych czatów

Dodano wyjaśnienie, dlaczego scroll myszą może nie działać w panelu terminala OpenCode, wraz z poprawką konfiguracji

What's new in 0.1.62

**macOS (Apple Silicon) support** — Llama Control now runs natively on M1 and newer Macs

**Chat from your phone** — turn on "Remote access" in Settings and pair over your home Wi-Fi with a QR code or PIN, nothing to install on the phone

**New Plugins tab**, starting with a JSON Extractor — paste text, define a schema (or have the model suggest one from the text), and get back structured JSON guaranteed to match it

**Connect a coding tool** — a new button in Chat generates a ready-to-paste OpenCode/Crush config pointing at your local server

A setup screen now appears on launch if llama.cpp isn't installed yet, instead of leaving the app looking broken

The sidebar can now be collapsed to a slim icon rail to reclaim screen space

HuggingFace search now loads more results as you scroll, and shows each model's added date

Added a refresh button to the saved chats list

Llama Control now runs on macOS (Apple Silicon, M1 and later) as well as Windows. A new Remote access option lets you chat from your phone over the same Wi-Fi network, pairing via QR code and PIN, nothing to install. The Plugins tab gains a JSON Extractor: paste text, define a schema, and get data back in exactly that shape, with llama.cpp guaranteeing valid JSON.

Closing the window now asks whether to hide to tray, quit and keep the model running, or quit and stop it, instead of silently killing a running model. A new tray icon offers quick Show and Quit completely actions. Find in page (Ctrl+F) lets you search the current view. The server flag editor accepts a pasted block of flags instead of adding them one by one. Dropping a folder onto the terminal area opens a new terminal there. Fixed Ctrl+V not pasting into the terminal.

Settings panel now shows the real driver version for AMD/Intel GPUs instead of always saying not detected. Chat and benchmark errors from the local server now show the actual reason (e.g. context size exceeded) instead of a bare HTTP 400. HuggingFace file browser marks quantizations you've already downloaded as Downloaded instead of showing an active Download button.

Fix: terminal panel would freeze silently when running full-screen TUI programs (e.g. opencode) in the packaged app. Root cause: Vite's production minifier mis-scoped a variable in xterm.js's terminal-mode-query handler, throwing an uncaught error the moment such a program probed terminal capabilities on startup. Production builds no longer minify the renderer bundle.

NSIS fix: desktop/start menu shortcuts now created on install. Safe PATH handling.

CLI command registration: llama-control is now available in terminal after installation. NSIS installer adds the app directory to user PATH automatically.

CLI command registration: llama-control is now available in terminal after installation. NSIS installer adds the app directory to user PATH automatically.

Terminals now run in their own process, so a console crash can no longer take the app down with it. Live statistics also follow the newest reading instead of the first one.

Each terminal session runs in a separate pty host process. Previously a fault in the Windows console host could take the whole app with it, closing the window without warning; now it costs only that terminal, which reports "session ended" while everything else keeps running.

Fixed a crash where programs running in the terminal (opencode and similar) believed they were talking to Windows Terminal, because the app was passing on environment variables describing the terminal it had itself been launched from.

Terminal output is batched before it reaches the screen, and xterm now draws through the GPU, so heavy output (build logs, full-screen tools) stays smooth.

Response time in the statistics is read from the server log, so it updates for traffic from any client, not only the app's own chat.

Throughput readings now prefer the newest measurement. A single early chat message used to pin the tiles for the rest of the session.

Image and audio categories removed from HuggingFace browsing, since the engine only runs text models.

Crash diagnostics: the app keeps a log in its data folder and records native faults, which is what made the terminal crash above possible to find.

Redesigned chat and Models tabs, a built-in split terminal, and accurate AMD/Intel GPU detection.

New chat bubble layout with streaming indicator and a collapsible chat/terminal history panel.

Built-in terminal alongside chat -- split into multiple panes, and it keeps running (along with whatever's inside it) even while you're on a different tab.

Models tab: cleaner detail cards, `.gguf` dropped from displayed names, correct Polish plural counts.

AMD/Intel GPU VRAM detection fixed -- now reads from the registry instead of a WMI field that silently overflows for cards >= 4 GB.

Quick Start shows the same startup progress modal as starting a model from its own page.

The biggest release so far. Three bugs that could lose your data or your settings are fixed, auto-tune finally does what it claims, and the interface got a visual overhaul.

---

Fixed: your chats were being saved empty

Every finished conversation was written to disk **without any messages**. Open an old chat from the sidebar and you saw a blank screen — the app was loading exactly what it had saved, which was nothing.

The reply was assembled from React state at a moment when that state was not yet guaranteed to be up to date, so what got saved was an empty list. Conversations are now assembled explicitly and saved complete.

**Chats you had before this release cannot be recovered — they were never actually stored.** From now on they persist properly, including across restarts.

Fixed: an interrupted write could erase everything

Your chats, your model tags and favourites, and your settings (models folder, per-model flags, saved profiles, encrypted API keys) each lived in a file that was rewritten from scratch on every change. If the app was closed, crashed, or the machine lost power partway through that write, the file was left damaged — and the app's response to a damaged file was to treat it as empty and overwrite it on the next save.

In other words: one badly timed interruption could wipe every conversation, every tag, or every setting you had.

Files are now written to a temporary file and swapped into place in a single step, so a reader always sees either the old file or the new one, never a half-written one. If a file is ever found damaged, it is set aside as `.corrupt-<timestamp>` instead of being discarded, so the data is still there to salvage.

Fixed: auto-tune

Three separate faults, each hiding the next:

**It crashed on its own first suggestion.** The AI proposed `--flash-attn` with no value; llama.cpp expects `on`, `off` or `auto` there, so the flag swallowed the next argument and the server died within a second. Flags that require a value are now filled in.

**It gave up after one attempt.** That failure was treated as an ordinary crash, and the recovery ladder can only adjust settings that are already present — the suggestion contained none of them. So the run ended immediately, without ever trying the configuration that was already working on your machine. Rejected command lines are now recognised as their own kind of failure, and the search falls back to your saved configuration instead of giving up.

**And when it did "succeed", it could make things worse.** Auto-tune never measured the configuration you already had. It compared the AI's suggestion only against variants of itself and saved the winner. Measured on a real model: the suggestion ran at 45 tokens/s against the saved configuration's 51, and the app replaced a carefully tuned setup with a generic one and called it an improvement.

Your existing configuration is now measured as one of the candidates and only replaced when something is clearly faster.

New: you decide when speed costs context

Ranking configurations purely on tokens per second is misleading. A setup with a small context is almost always faster than one with a large context, simply because there is less to keep in memory — so "fastest wins" quietly throws away the context length you chose.

When auto-tune finds something faster that would shrink your context, it now **saves nothing**. It shows you both figures — for example *"faster (72 t/s vs 51 t/s), but the context drops from 128,000 tokens to the small default"* — and lets you choose. Your own configuration is restored in the meantime, so the app is left working either way.

Also in this release

**Model vendor icons.** Qwen, Llama, Mistral, Gemma, DeepSeek and Nemotron models are now recognised by filename and shown with their logo in the model list and detail view.

**Refreshed interface.** A dark sidebar carrying the artwork, a proper wordmark, and a tidier title bar.

**Long model names are readable again.** Names that differ only at the end (…Q4_K_M vs …Q4_K_XL) wrap onto two lines instead of being cut off identically.

**Cancelling auto-tune responds immediately** instead of after up to 90 seconds, and a failed attempt no longer stalls for the full timeout.

**Dates follow the interface language** instead of always using Polish month names.

---

Windows may warn about an unsigned installer — see the README for how to get past it.

A maintenance release that mostly makes the app a better citizen on your machine.

**Checks for updates far less often.** The app used to ask GitHub for a new version every 5 minutes. It now checks shortly after launch and then every 6 hours, which is plenty for something that ships a few times a week — and it stops the app quietly generating hundreds of requests a day in the background.

**The installer now has a stable filename.** It is always `Llama-Control-Setup.exe` rather than carrying the version number. Updates are unaffected.

Everything from 0.1.50 (English interface outside Poland, security hardening) is included.

Security hardening, an interface that speaks your language, and a new home for the project.

**The app now starts in English outside Poland.** It picks its language from your Windows locale on first run instead of always defaulting to Polish. You can still switch it manually in Settings at any time. Dates follow the chosen language too.

**Hardened the app window.** Added a Content Security Policy, restricted outbound links to real web addresses only, and blocked page navigation and popups inside the app window.

**The project moved to github.com/Mil0s7/llama-control.** Existing installs keep updating normally; this build points at the new address directly.

Windows may warn about an unsigned installer — see the README for how to get past it.

Redesigned model detail view: new model info card with a computed KV cache size estimate, server card with save/clear log buttons and a 3-column parameter layout, live CPU/RAM statistics with a sparkline chart plus TPS/prompt-speed/response-time/uptime readouts, and a bottom actions bar (start/stop, edit parameters, duplicate, export). Also fixes the sidebar artwork rendering inconsistently across window sizes.

Fix: sidebar artwork now shows the full llama image without cropping (object-contain), matching the intended layout at any window size, at 70% opacity.

Added the llama-on-mountain artwork to the bottom of the sidebar, matching the requested mockup layout.

Fixed stale llama.cpp update message, added auto-dismiss to the update toast, replaced native confirm() dialogs with a styled modal, fixed the chat empty-state message when the server is off, and reordered/flagged the language switcher (English first with US flag, Polski second with PL flag).

Added real per-slot progress bars to Benchmark (exact % for quality suite and perplexity, honest active/queued indicator for speed and custom-prompt modes) instead of plain status text.

Benchmark now supports an open-ended list of models to compare instead of a fixed Model A/Model B -- each slot can also independently pick a saved profile, so you can compare different models or the same model under different flag configurations in one run.

Added perplexity benchmarking against the real wikitext-2 test corpus (the same reference llama.cpp's own examples use) via llama-perplexity -- a recognized, directly comparable metric, not a private benchmark.

Benchmark now stops the server after a run finishes instead of leaving it loaded, and shows a live server log so you can confirm the model actually started up correctly.

Rebuilt Benchmark around real methodology: llama-bench for speed (statistically repeated pp/tg throughput, GPU throttle detection), an 18-task deterministically-graded quality suite (math/logic/code/format/knowledge/traps, ported from a proven external benchmark script), and the original free-text prompt mode kept for ad-hoc testing.

Benchmark can now judge response quality automatically using the AI provider configured in Settings (Claude/GPT/Gemini) -- a 1-10 score and short verdict per model, not just speed.

Benchmark now has a test-suite mode: 4 built-in tasks (logic/math, code, reasoning, instructions) run automatically on each model so you can compare quality, not just speed.

Fixed a false "server stopped unexpectedly" notification that could appear when switching between models (e.g. during a Benchmark comparison).

New Benchmark tab for testing/comparing models, category filter removed from Models (HuggingFace only), and real Claude/OpenAI/Gemini logos in Settings.

The model-type category menu is now a dedicated sidebar column (matching Chat/Models layout) in both HuggingFace and the local Models tab.

Added a model-type category filter to the local Models list, matching the same category menu used in HuggingFace search.

Full PL/EN interface translation with a language switcher in Settings, and a smaller/less prominent New Chat button in the Chat tab.