Matilda Jev
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity5.4/10
Open Apache-2.0 weights (~26.1B, ~49 GiB bf16) with a bundled /v1/systemone server, OpenAPI docs at /docs and custom AutoClasses. Docs are the HF card. No third-party integrations, hosted API or SLA.
- Capability6.5/10
Choice / noul / score with a 255-option readout, a stored calibration temperature (no ECE published) and optional images. No latency figure on the current card. Decision Index 59.59 is Maincode's own full run, not on the official board.
- Adoption5/100
Two days old: 6 HF likes and ~180 downloads (2 Oct). Maincode is an Australian lab with an existing Matilda model line, but no social spike, integration or production use found for this decision model yet.
Vendor claims
- Decision Index 0.2.1 = 59.59 (raw 69.18, breadth 58.39) over 150,317 requests for the v1.3 weights, 'currently awaiting official submission'; the 30 Sep v1 card said 59.26[Vendor claim — not independently verified]
- Area scores (chance-corrected skill × 100): Knowledge & Reasoning 44.09, Language 66.27, Retrieval & Classification 61.07, Tools & Automation 79.07, Arts & Human Taste 43.85[Vendor claim — not independently verified]
- Validation on AMD MI355X: 25/25 smoke-test questions passed; 23 requests matched the original checkpoint's probabilities exactly[Vendor claim — not independently verified]
Matilda Jev (styled MATILDA-jev on the card) is Maincode’s open-weights System One model: a ~26.1B backbone with a 255-option readout that returns a probability for each option of a choice, noul (yes/no) or ordered score question. It takes text or JSON state and optional images. The HF card went up on 30 September 2026 (08:21 UTC); Maincode pushed v1.3 weights on 2 October at 00:01 UTC (21:01 BRT on 1 Oct). The license is Apache-2.0, and the repo now ships the LICENSE file.
Specs
| Attribute | Value |
|---|---|
| Developer | Maincode (Australia) |
| Params | ~26.1B; ~49 GiB in bf16 before runtime overhead |
| Readout | 255-option decision readout (readout.safetensors) + stored calibration temperature |
| Types | choice, noul, score |
| Inputs | text or JSON state, optional images |
| Serving | bundled maincode_jev_serve → POST /v1/systemone, OpenAPI at /docs; or AutoClasses with trust_remote_code=True |
| Tested on | AMD MI355X (other accelerators untested for this build) |
| License | Apache-2.0 |
| Released | 30 Sep 2026 (HF); v1.3 weights 2 Oct 2026 UTC |
Decision Index claim
Maincode reports a full Decision Index 0.2.1 run (150,317 requests) for the v1.3 weights:
| Model | Params | Index | Raw | Breadth |
|---|---|---|---|---|
| Matilda Jev v1.3 (Maincode run) | 26.1B | 59.59 | 69.18 | 58.39 |
| Matilda Jev v1 (30 Sep card) | 26.1B | 59.26 | 68.89 | 58.06 |
| Jev 1.13 (official board) | – | 57.91 | – | – |
UnverifiedClaim. The card says the model is “currently awaiting official submission”. The official Decision Index board had no Matilda row at our 2 Oct check (its data is from 28 Sep). By area, Matilda is strongest on Tools & Automation (79.07) and weakest on Knowledge & Reasoning (44.09). The card no longer publishes a latency figure, and the stored temperature comes without an ECE or method write-up.
Fit / anti-fit
Fit: a large, self-hosted, Apache-2.0 decision model with images and a Jev-compatible local endpoint, for teams with a ~50 GB GPU.
Anti-fit: anything that needs board-verified accuracy, published latency or calibration numbers, or a hosted API with support.
Why it is in the catalog
Trained decision weights plus a typed probability API in the System One shape, publicly downloadable under Apache-2.0. Same bar as the other open peers. The Index numbers stay UnverifiedClaim until the board or an independent run lands.