Skip to content

Releases

v1.0.0-pre.4

v1.0.0-pre.4 Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 11 Nov 07:55

A fourth pre-release before 1.0.0. It is a pre-release, so the update check and the website stay on 0.9.9. Download an installer below to try it.

  • An MLX model can be the delegator on a Mac. It routes each message to an expert and decides whether to answer or build, as a GGUF does, in a helper of its own on mlx-lm. Pick one in Settings, General, where the delegator's menu now offers MLX models wherever MLX runs.
  • CI runs MLX on a Mac, against a small real model, so the helper is checked on every push.

Everything in pre.3 is here too. The rest is in the CHANGELOG.

v1.0.0-pre.3

v1.0.0-pre.3 Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 11 Nov 07:37

The last of three pre-releases before 1.0.0. It is a pre-release, so the update check and the website stay on 0.9.9. Download an installer below to try it.

  • The server speaks HTTPS when asked, with your own certificate or one Crucible makes for this machine and keeps. Settings shows its SHA-256 to check it by.
  • Builds on LFM2 models write their files. Their tool calls were taken for talk.
  • Closing Crucible while the server or Bench is answering no longer risks a crash.
  • gpt-oss chats past the first turn show the model its earlier answers in its own format.
  • Two GPUs of the same model are both used. The second was left out.
  • A Mac counts cached files as available memory, as Activity Monitor does, so models that fit are no longer let go or refused.
  • A message about your code is answered rather than built, as in "Code in main.py crashes".
  • The server's tool calls and refusals behave as OpenAI's clients expect. A required call comes back alone, and a template that refuses a conversation says why.
  • Preview: a page opened by its address can no longer overwrite a project page's saved data, and More in the Tune panel works again.

Everything else is in the CHANGELOG.

v1.0.0-pre.2

v1.0.0-pre.2 Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 11 Nov 02:13

The second of three pre-releases before 1.0.0. It is a pre-release, so the update check and the website stay on 0.9.9. Download an installer below to try it.

  • The server takes tools, pictures, JSON and embeddings, the way OpenAI's clients expect, and listens on IPv6 too. A provider's model can be served, to other computers only with a key.
  • Builds send their tasks to the right expert. With LFM2.5 as the delegator, 27 of 31 test tasks go where they belong, up from 19: styling work no longer lands on Writing.
  • A better delegator, suggested. Where your cards hold it, Settings suggests Qwen3 4B, which routed every one of the nine-expert test.
  • A made app reloads on Windows and macOS without going blank.
  • Dedicated VRAM only behaves on Windows and macOS, and the delegator makes room beside an idle MLX model on a Mac.
  • Model cards' own pictures show in Models up to a megabyte.

Everything else is in the CHANGELOG.

v1.0.0-pre.1

v1.0.0-pre.1 Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 11 Nov 00:11

The first of three pre-releases before 1.0.0. It is a pre-release, so the update check and the website stay on 0.9.9. Download an installer below to try it.

  • Preview pages keep what they save across a restart, and a made app reloads on Linux instead of going blank.
  • Bench vs base finds the base from the recipe rather than guessing it from file names, and offers Get base when it is missing.
  • Models shows a repository's own picture when its card names one.
  • Windows fills the fastest card first, reading each card's PCIe link as Linux does, and a Mac loads a cold model faster.
  • Pictures are read whole by models that need it (Gemma 3 and 4), and a vision projector goes where it fits, measured.
  • A chat turn out of tool rounds answers from what it found, where a small model could end with no answer at all.
  • The Bench gives the same answers run after run on the same machine.
  • Gemma 4's thinking goes in the thinking fold instead of opening every reply.
  • Releases now take minutes instead of hours, so fixes reach you sooner.

Everything else is in the CHANGELOG.

v0.9.9

Choose a tag to compare

@github-actions github-actions released this 07 Nov 03:09

Faster across the board, and steadier.

  • Models stay loaded while they fit. A follow-up question loads nothing, and going back to an expert that is still in memory is instant. Five routed prompts from a cold start took 160 s, down from 428.
  • Cold loads are about twice as fast: Qwen3.8-27B in 35 s instead of 73, flint in 40 instead of 78. On Linux the fastest card fills first.
  • The Bench asks many questions at once. The small LFM2.5 model's Quick run takes under 3 minutes instead of 13.
  • Fine-tuning is three to four times faster, and every fine-tune now ends at a GGUF. Before, on a machine with a downloaded runtime, it ended at a folder Crucible could not load. New recipes learn at 1e-4.
  • Long conversations open at once, replies stream without the window falling behind, and typing keeps up beside a long chat.
  • The delegator on a Mac: MLX models are listed, and Get one fetches a small GGUF delegator when the folder has none.
  • Steadier: a killed orchestrator, a model with no vocabulary, a Gemma picture with a large Batch and a conversation full of pictures no longer crash or stall, and crash.log says why llama.cpp stopped.

Everything else is in the CHANGELOG.

v0.9.8

Choose a tag to compare

@github-actions github-actions released this 27 Oct 07:56

A hotfix: a vision model no longer crashes Crucible when it finishes loading.

  • Qwen3.8, flint and other models that read pictures carry a projector beside them. It went on the first card whatever was on it, and on a full card llama.cpp ended the whole program at 100%. It now goes on the card with the most room, or on the processor when no card has room, where pictures are read more slowly.
  • The server shows its IP and port, the address another computer on your network types in to reach it.

Everything else is in the CHANGELOG.

v0.9.6

Choose a tag to compare

@github-actions github-actions released this 27 Oct 05:10

One box for asking and building, a model served on your network, and a Bench that opens on everybody's results.

  • Ask or build, in one box. Build is no longer a tab. Ask a question and the right expert answers it. Ask for software, like "a budgeting app that charts my gross and net revenue", and it becomes a build, with its plan, agents and check in the conversation. Code, Source and Preview open above it once there are files.
  • Serve a model on your network, the way LM Studio does. Settings, Server: the delegator or any model here, a port, this machine or the local network, and an optional API key. Any OpenAI client can talk to it.
  • The Bench opens on the public board, with the best models first, each with a Download that opens it in Models, then the spider chart and every result. Add benchmark in the corner runs your own, and Bench vs base puts a fine-tune beside the model it came from.
  • Logos in Models, the original maker's for a repack.
  • A model larger than the card runs partly on the processor, where before it failed to load.
  • Builds get the context the cards can hold, route their tasks to the right expert more often, and add an expert on a model that fits.
  • History moved into the right-hand panel.

Everything else is in the CHANGELOG.

v0.9.5

Choose a tag to compare

@github-actions github-actions released this 26 Oct 18:39

Benchmarks with a public board, models condensed to fit, and an editor in Build.

  • The Bench tab runs local models through nine benchmarks in seven skills: MMLU-Pro, ARC-Challenge, GSM8K, MATH-500, HumanEval, MBPP, IFEval, HellaSwag and TruthfulQA. Every Crucible asks the same items the same way, so a score from one machine stands beside a score from another, and a model that thinks first has room to.
  • A spider chart compares up to three results, with the numbers under it, and a result opens to each benchmark with its margin of error and the answers that were wrong.
  • The public board, mattsaund/crucible-bench. Post a result and it goes up from your GitHub account, checked against Hugging Face, and sits beside everybody else's in the tab. A result is names and numbers, nothing typed.
  • Condense a model from Create: a Q8 that does not fit becomes a Q4 that does, and the largest models go down to IQ2 or IQ1 after a pass over sample text finds the weights that matter most. Every size is measured first and marked by how it fits.
  • An editor in Build. Tabs, a file tree, find and replace, search across the project, Ctrl+P to open a file, and a terminal under it.
  • A fine-tune is scored on data it never trained on, before and after.
  • Quicker where it counts. The delegator decides in a quarter of the time, models like gpt-oss fit a longer context on the same card, and a model split across cards holds back less of each.

Everything else is in the CHANGELOG.

v0.9.4

Choose a tag to compare

@github-actions github-actions released this 24 Oct 03:24

Models from Hugging Face in the window, the newest llama.cpp, math in replies, and the system's font.

  • The Models tab searches Hugging Face for GGUF and MLX models, says whether each quantization fits on this machine's cards, runs partly on the CPU or is too large, and downloads into the models folder with a progress bar. A download carries on after a restart and is checked against Hugging Face's SHA-256.
  • llama.cpp b11846, which runs flint and the other newest models. A model llama.cpp cannot run now says so, instead of "Too large for the cards?".
  • A model that fits on the cards loads. flint (33 GB) was refused on three cards with 36 GB free; before refusing, Crucible now asks llama.cpp what the load would really take.
  • Math is typeset in replies, from LaTeX, and code blocks show the language and a Copy button, without line numbers.
  • The system's font, or any installed font from Settings, General.
  • Picking an expert in the box loads it at once, and keeps it loaded while it is picked.
  • The chat box holds still while typing.
  • Fixes from a review of everything since 0.8.6: a fence like ```c++ froze the window, a page in the preview could run script in Crucible, runtimes never downloaded on a Mac, RTX 20 and GTX 16 cards failed with the CUDA runtime, closing the window could leave a download running, and more.

Everything else is in the CHANGELOG.

v0.9.3

Choose a tag to compare

@github-actions github-actions released this 21 Oct 22:54

The first start after an update runs models again.

  • Runtimes load on the first start after an update. Updating from 0.9.1, Crucible could not load any runtime until it was restarted, so no model ran. A runtime built against another llama.cpp now stays closed until its replacement is in.
  • Running the installer again over a current install no longer warns that the runtimes and the fine-tuner are missing.

Everything else is in the CHANGELOG.