What’s new in NEURAX

v1.1.1

October 2026

Linux: NEURAX 1.1.0, now available as desktop installers

• Brings the Studio and runtime updates from NEURAX 1.1.0 to Linux desktop installers.
• Packages are available as .deb, .rpm, and .AppImage, each with a SHA-256 checksum.
• Linux only: this release does not include macOS or Windows installers.

v1.0.1

October 2026

Linux: the studio stays open when semantic memory is unavailable

• Fixes a desktop crash when sending a prompt on Linux: without the optional ONNX Runtime library, the engineer falls back to lexical memory search and the studio stays open.
• Semantic memory search stays unavailable on a fresh installation; the engineer and lexical memory work as before.
• Linux only: macOS and Windows are coming soon.

v1.0.0

October 2026

A trained model ready to serve, from data prepared as an engineer would

• Tables are read as the profile decides: categories one-hot, identifiers and constant columns left out, possible leaks pointed out.
• A test part is held out by default, measured on the delivered checkpoint; early stopping is on by default.
• The export carries serve.py: an HTTP API for the model, straight from its folder or its GitHub repository, exported as one commit with its weights.
• Runs live in Documents/NEURAX and open in the file manager; saving large weights no longer passes through the page.
• Five findings of the v0.25.0 security audit are closed.

v0.25.0

October 2026

A new Engineer, in Rust, that works in the Studio like a person

• The agent works only through the Studio's own controls and reads every result back from the screen before its next step.
• After every change NEURAX compiles again; the agent reads the newest diagnostics and does not stop while the design is incomplete.
• Once the client approves in Decision it resumes by itself, follows the training to its end, corrects a run that falls short, and finishes the request: weights in the client's folder, model code, targets checked.
• French and English requests are understood however they are written, and every request ends with two next steps to click.
• Every block parameter is editable with a real value; data and compute come from this machine only.

v0.24.7

October 2026

The Engineer verifies a design before asking for approval

• The agent checks the requested design against a fresh compiler result before opening Decision.
• It keeps repairing while compiler and machine evidence point to a better result, then explains any blocker it cannot resolve.
• The detected machine profile is applied before the proposal and reused for training after the client approves.
• Provider failures remain visible as errors instead of being reported as successful completion.

v0.24.6

October 2026

The Engineer chat keeps each request in its own context

• Requests are sent immediately with the machine profile already detected; a new measurement is optional.
• A rapid double click or repeated Enter sends one request, not two.
• Each chat keeps its own requirements; a new chat starts a new conversation.
• Explicit size, family, training and Time Machine requests are recovered when the provider misreads them, then checked against the client’s words.

v0.24.2

October 2026

A full disk no longer loses a run

• A full disk interrupts a run, which then resumes with its last validation.

v0.24.1

October 2026

The training runtime matches the tested versions

• The training runtime installs the versions the test harness is built on.
• Text and tables are read under pandas 3, which that runtime installs.

v0.24.0

October 2026

Fine-tune from an earlier run

• A fine-tuning run can start from an earlier NEURAX run’s export.
• A warm start reads the data exactly as the run it starts from did.
• A template loads with its own values and is analysed at once; a slider’s bounds never refuse a published template.
• A small model’s step time includes the framework’s own cost, measured on the machine.
• Decision states a duration NEURAX chose as the most a run takes.

v0.23.0

October 2026

Reading the machine without blocking it

• Loading a template keeps what you said about your machine.
• The compiler refuses an output that still declares a template’s count, and names the fix; one value under two names with two numbers is said out loud.
• Memory, disk and processor are read on macOS and Windows, and reading the machine never holds the server.
• Text is encoded once and a dataset is read and split once per run.
• Runtime setup streams its progress, and provider keys are protected.
• A design is calibrated by the runs of its own scale.

v0.22.4

October 2026

GitHub export, clearer provider errors

• Export creates GitHub repositories from suggested names and says when the token is not ready.
• The assistant explains provider 503 retries.
• Image class outputs are checked before training; Python is detected and child processes are reconciled.

v0.22.3

October 2026

Hardening, and a private training runtime

• Provider keys stay in memory only; an unused JWT dependency was removed and frontend advisories patched.
• Training installs and verifies a private PyTorch runtime, and explains a missing runtime before you approve.
• The current design is recovered after a refresh; design files are JSON or TOML only.
• Padding masks are preserved across token streams, and image dataset counts match the training loader.
• Python is found through the Windows launcher.

v0.22.2

September 2026

The Engineer takes blocks from every palette

• The Engineer adds blocks from every palette, decides by itself how to start, and stops only for the training approval.
• It says so before removing a block no palette can add back.
• It slims a heavy model by its blocks, and never misleads about it.

v0.22.1

September 2026

The assistant starts on Linux

• The desktop app finds the assistant where the Linux packages put it.
• Theme documentation.

v0.22.0

September 2026

The NEURAX Engineer, and the first public release

• The NEURAX Engineer designs, repairs and runs through a request on any provider.
• Time Machine projects training year by year from real compilations.
• Decision proposes the budget at the top of the cost range.
• Each machine is calibrated on its own finished runs, and a run’s length comes from the model and the data.
• A trained model is delivered with a README that says how to load it; designs are JSON or TOML files.
• AWS compute connections are verified through STS-assumed roles.

v0.21.0

September 2026

An assistant that acts on what it sees

• The assistant requires a fresh look at the screen before it acts, and honours the architecture you name.
• It verifies completion against your request, and autonomous mode is now the default.
• Each desktop session runs the assistant on its own private port.

v0.20.2

September 2026

Windows builds, and nothing names the private repository

• Windows builds were fixed, and nothing shipped names the private source repository.
• Signing out revokes the session on the server.

v0.20.1

September 2026

Signing in no longer stops the studio

• Signing in no longer stops the studio from compiling.
• Plan lookup, the real account display and a hard gate were fixed, along with a token that appeared in a URL.

v0.20.0

September 2026

A training step is not a forward pass

• The cost model now counts the backward pass: training time, cost, energy and carbon were built on a forward pass alone.
• Gradient checkpointing is charged, replicas are counted, and carbon intensity is no longer one constant for the whole world.
• Routes that run your Python, read the paths it names or hold a memory store now require a per-process token over loopback.
• Decision is two pages, with a budget NEURAX proposes and one sign-off that will not record a decision without its reason.
• Fine-tuning workspace preview, and ModelScope imports.

v0.19.2

September 2026

Desktop icon and sharper logos

• The application icon is installed, and the logo renders sharply on high-density screens.

v0.19.1

September 2026

Keep the studio open after sign out

• Signing out no longer closes the studio.

v0.19.0

September 2026

Studio, metrics, themes and agent computer use

• A new studio with metrics and selectable themes.
• The assistant can operate the studio itself.
• The compiler refuses what cannot be counted, named or told apart.

v0.18.0

September 2026

One block, a fragment, or a whole model

• A new compiler lowers every block to primitive operations and prices one block exactly as it prices a model.
• A design with one broken block keeps every other block's figures, and names what failed with its own reason.
• Every figure carries its basis (exact, measured or estimated), and what is not known is said as unknown.
• Sweeps across batch sizes, precisions and device counts.
• Six architecture families end to end; GAN and recurrent networks were removed.

v0.17.0

September 2026

From a prompt to a trained model

• The assistant builds real models from a prompt, on your own hardware.
• Datasets are profiled where they are: on this machine or in a bucket.
• Training reports only real runs, and labels examples as examples.
• Decision now comes before Training, so the person who signs sees the numbers first.
• A missing Python package is reported as something to install, not a server error.

v0.16.0

September 2026

NEURAX runs the designs it costs

• One gate every model passes before it is trained: generated code must match the approved design.
• Read and edit the generated project beside the running training.
• The hardware probe reports what the machine can run, not just what it contains.
• Choose where the data is and where the run happens.
• The Production workspace was removed; Decision is the workspace for the person who signs.