ChaosEngineAI v0.9.8

What's new in v0.9.8

The first release since v0.9.4. Versions 0.9.5 to 0.9.7 were internal builds, so everything below is new to anyone on 0.9.4 or earlier.

Highlights

• DFlash on Apple Silicon finally runs. Since dflash-mlx 0.1.5, every DFlash reply had quietly fallen back to normal decoding while the load note still said DFlash was active. Fixed: Qwen3-4B now writes at 113 to 138 tokens per second against 53 without it, with identical output.
• Newer MLX underneath. Releases now ship mlx 0.32, mlx-lm 0.32 and transformers 5.20. Setup installs resolve mlx-vlm 0.7.10 and diffusers 0.41. Current Gemma 4 conversions load (the old mlx-vlm could not read them), and every release build pins the exact versions it was tested with.
• Images in chat on Apple Silicon. Vision models loaded through mlx-vlm (Gemma 4) now see attached images. Before, the image was always dropped.
• Models stop loading when you ask. Unload cancels a load in flight, and picking a different model replaces the one that is loading instead of queueing behind it. The Dashboard lists leftover model processes with a Stop button each and a Stop idle button, so a stuck worker can no longer hold 20 GB of memory with no way out. Send is disabled while a model loads, so a message can no longer go to the previous model.
• Honest memory and speed estimates. The launch screen works out the real KV-cache size for each model's architecture and estimates speed from your own past runs or your machine's memory bandwidth. It warns before you load a chat model that may not fit. Image and video estimates know which engine they run on and count a chat model that is already loaded.

New models

Chat

• Kolibri 1 (Aleph Alpha, Apache 2.0): a 78B mixture-of-experts model with 3.5B active per token, built for German and English, with a 256K context.
  • Runs on Apple Silicon through community MLX builds (4-bit, 3-bit, 2-bit) at about 60 tokens per second (3-bit, M4 Max).
  • A new llama.cpp Q3_K_M build (37.6 GB) also runs on Windows and Linux.
  • The Thinking control sets its reasoning level.
• LFM2.5 24B-A2B (Liquid AI): MLX 4-bit and 8-bit, and a GGUF build. About 100 tokens per second (MLX 4-bit) and 140 (GGUF) on an M4 Max.
• Flint (Recursive, Apache 2.0): a 39.8B coding-agent mixture-of-experts model with 2.5B active per token and a 256K context. Runs through llama.cpp: the official INT4 GGUF (26 GB, Apple Silicon with 48 GB or more) and a community Q4_K_M for Windows and Linux. About 112 tokens per second on an M4 Max, with reasoning and tool calls working.
• Muse Glimmer 30B GGUF, with image input and its DFlash 2 drafter: 29 tokens per second against 19.5 without the drafter.
• Qwen3.8 Flash Next GGUF (111 GB, for machines with 128 GB or more) and Qwen3.6 35B-A3B with MTP through TensorFold.
• Also new since 0.9.4: Kimi K2.6, Llama 4 Scout, MiniMax M3, Mistral Large 3, Ornith 1, Qwen 3.8, Bonsai (1-bit and ternary builds that fit a 27B model in 4 to 9 GB), Nemotron 3.5 Lightning and MiniCPM 5.

Image

• Qwen-Image-2.1 Turbo through mflux: 1024x1024 in about 95 seconds on an M4 Max (8 steps).
• FLUX 3 Image Klein (distilled and base) through mflux: 1024x1024 in about 83 seconds on an M4 Max (16 steps, 23.5 GB peak). Black Forest Labs' licence must be accepted on Hugging Face for each repo, and your token needs read access to gated repos.
• Qwen-Image-2.1 (7B) on Apple Silicon through MLX, with a stable-diffusion.cpp variant for other platforms. Qwen-Image-2.1 and its Turbo are under the Qwen Research licence (non-commercial).
• FLUX.2 klein 4B, plus Qwen-Image, Z-Image Turbo and FLUX.1 through the diffusers lane.

Faster generation

• DFlash (Apple Silicon) fixed, as above.
• Thinking Off now works for Qwen 3 / 3.5 / 3.6 / 3.8 on Apple Silicon, and the switch is sent to llama.cpp models too. Before, they reasoned anyway and spent the reply budget on it.
• TensorFold 0.6.6 (Apple Silicon): exact speculative decoding with no change in output. Install it from the launch settings.
• llama.cpp b11810 bundled: DFlash 2, EAGLE-3 and draft-free ngram-mod, picked automatically from the drafter files next to a GGUF.
• MTPLX 2.12: reasoning now streams correctly.
• Hybrid models (Qwen 3.5 / 3.6 / 3.8, Qwen3-Next, Nemotron-H, Mamba) reuse the chat history between turns instead of re-reading the whole conversation.
• TurboQuant is labelled a memory saver, not a speed option.

Image and video

• mflux engine (Apple Silicon) now at 0.23, in its own isolated environment. The Image Studio banner offers an update when an older version is installed.
• stable-diffusion.cpp lane: FLUX.1 schnell and dev fetch their text-encoder and VAE files on first use, and image runs show live progress.
• Wan videos no longer come out burnt and over-saturated on Apple Silicon.
• Catalog LoRA variants (Hyper-SD, Turbo-Alpha, CausVid) now download their LoRA on first use.
• Image and video Discover list every variant once. Variants that share a download are marked.

Voice I/O (in progress)

Speech-to-text and text-to-speech are not in this release. The backend is in place (Whisper and Parakeet on Apple Silicon, faster-whisper on Windows and Linux, Kokoro for speech), but the Voice screens stay hidden until they pass testing in the packaged app. They are planned for a following release, with voice input and read-aloud built into Chat.

Chat and Settings

• Chat > My Models lists chat models only. 3D tool parts, text encoders, voice models and bare weight files no longer show up as chat models.
• Settings > Diagnostics is split into three tabs: Diagnostics, Accelerators and Disk usage.
• Disk usage rows expand and can be deleted. The shared Hugging Face blobs store is now broken down by repo, and any cached repo can be deleted from there; blobs another repo still uses are kept.
• Streaming follows the bottom. In Chat, Compare and the HTML challenge the view now stays with the newest text until you scroll up. Before, one long chunk made it stop following, and Latest had to be pressed again and again.
• Multi-quant GGUF repos show the size of the one file that loads, not the whole folder, in the library and the memory bar. A model that fits no longer looks too big to run.

Fixes

• GGUF chat: models that reason even with reasoning off (Muse Glimmer) showed an empty reply. Launching a catalog model from Discover skipped its chosen quant and its drafter file.
• Apple Silicon chat: LFM2.5's reasoning leaked into the answer. A reply cut short (for example by the repeat guard) made the next action fail with "invalid result payload".
• Downloads that would not load: a cancelled download of another quant in the same repo, a library scan taken mid-download, or an organisation name ending in "models" each blocked a finished model.
• Linux AppImage stuck on "Still loading". Fixed; the log path now appears on the loading screen. Not yet tested on Fedora.
• Local GGUF folders from ModelScope and symlinked model folders are now listed.
• GGUF downloads fetch only the quant a catalog row names (one 16 GB row used to pull 68 GB).
• Dead catalog links repaired; an audit script checks every row.
• Release builds include the dflash package, the Setup installers and a pinned llama.cpp build, so GGUF models work out of the box.
• Tool use on Apple Silicon no longer loops.
• Loading a model no longer hangs forever when a second load is requested: repeated requests for the same model share one load.
• My Models and the chat picker no longer come up empty while the first library scan is still running.
• MTPLX is no longer chosen for models without an MTP head, and the TensorFold toggle names DFlash 2 and MTP drafting. A cold TensorFold start says it can take a few minutes.
• Extras overlay: an older huggingface-hub or transformers left by a Setup install can no longer shadow the ones the app ships.
• MLX probe: one stray kill of the startup check at launch no longer turns MLX off until you restart; it is retried once. The launch screen's context setting is capped at the model's own window, so a stored 262K no longer triggers a false memory warning on a 32K model.
• Hub listings work on Python installs without a certificate bundle (Discover came back empty there).
• Downloads skip weight shards that no index file uses, and the Gemma 4 entries are merged into one family.
• Image and video refuse a model larger than free memory before loading it, and unload idle models first so the check sees real headroom.
• Models that only mlx-vlm can load (Flint, Qwen3.8 Flash Next, GLM-5.3 Flash, DeepSeek-V4, Bonsai) now load through it when mlx-lm lacks the architecture. Flint's official MLX build still cannot load; use the GGUF.
• The memory check counts memory macOS can reclaim, so large models are no longer refused with plenty free.

Removed

• The DeepSeek-V4 8-bit, BF16 and Pro MLX rows, which nothing in the app could load. The 4-bit row stays and runs through TensorFold.
• The DDTree tree-budget slider on Apple Silicon. Its tree verification was never finished; DFlash runs instead.

Known limits

• stable-diffusion.cpp and llama-server-turbo are not bundled in the release build. The sd.cpp image variants need the sd binary built from a source checkout.
• MLX models that are multimodal but load as text only (Qwen 3.5 / 3.6 / 3.8 on MLX) do not see images yet.
• Not tested on real hardware: the Linux fix on Fedora, sd.cpp FLUX on CUDA, the CausVid LoRA on a real Wan pipeline, Qwen3.8 Flash Next GGUF (needs more memory than the test machine has).
• The CausVid LoRA is CC BY-NC; Qwen-Image-2.1 and its Turbo are non-commercial.
• Kolibri 1 chat history is re-read each turn once it passes about 500 tokens.

Downloads

Platform	File
macOS (Apple Silicon)	ChaosEngineAI_0.9.8_aarch64.dmg
Windows	ChaosEngineAI_0.9.8_x64-setup.exe
Linux	ChaosEngineAI_0.9.8_amd64.AppImage, ChaosEngineAI_0.9.8_amd64.deb

The in-app updater reads latest.json; the .sig files and ChaosEngineAI_aarch64.app.tar.gz next to the installers are its inputs.

Upgrade notes

• If your install keeps offering v0.9.4 although you are on it: v0.9.4 was built as 0.9.3 by mistake. Installing this release ends it.
• No breaking changes: settings, sessions and APIs carry forward.
• On first start the app moves older copies of mlx and mlx-lm that earlier Setup installs left in its extras folder to extras/<python>/stale-core/, so they cannot shadow the newer versions it ships. If the video runtime or DFlash then misbehave, reinstall them from Setup.
• TensorFold, MTPLX and mflux each install into their own environment under ~/.chaosengine/ on first use. TensorFold and mflux need Python 3.11 to 3.13 on the Mac.