[BlivionIaG]

Change

Admit official LiquidAI LFM3-1.7B and LFM3-VL-2B MLX 4/6/8-bit metadata, without registering inference. Validate family geometry, quantization layouts and tensor headers. Add a host inspector and rejection tests through the normal build graph.

Refs #1491. This is the admission foundation requested in #1502; text inference in #1514 builds on this, with VL #1502 stacked above text. The original admission commit and authorship are preserved. CUDA follow-up #1508 already carries a cherry-pick of that admission.

Code pointers

• Family geometry and quantization validation.
• Tensor admission.
• Checkpoint inspector.
• Host tests and build registration.

Verification

Pinned base: b30d5b9d3. Zig 0.17.0, macOS 26.5.1, M4 Pro 48 GB: zig build test-lfm3-host tf-lfm3-check -Dcpu=apple_m1 passed 8/8 steps and 13/13 host tests; the admission Python suite passed 12/12. This slice performs no GPU inference and makes no runtime or performance qualification claim. The recipe links the weights' LFM Open License v1.0.

ATC overlaps from the full implementation include conflicts with concurrent work; this smaller admission slice is intended to establish shared ancestry. Integration must be reconciled before landing.

Admitted checkpoint revisions

This slice admits metadata and tensor headers only: no engine registration, model execution or inference.

• LFM3-1.7B MLX 4-bit: d0ca73dcfa70ce18733754a492cdc1fb0123262d.
• LFM3-1.7B MLX 6-bit: 26be898c36b545d625d84ac0761c6c3eb664aef3.
• LFM3-1.7B MLX 8-bit: ee7667f3cfccb2a41725f91820531823fcabc6b7.
• LFM3-VL-2B MLX 4-bit: c823fade98e35a3d6942ad4d38b1ea15f675cd12.
• LFM3-VL-2B MLX 6-bit: 6d71bfae121abedc75f6d531fc7f6ceb4f5e3501.
• LFM3-VL-2B MLX 8-bit: 83336d5be3874a2bdd3297bde77b2114dbd3a816.

Dependents: text Metal #1514 and CUDA #1508. The shared-name extraction suggested for XPU is under review; it is not included in this head yet.

[ashhart]

Thanks for splitting the admission out. It reads well and the host tests cover it. Could you add a short PR description: what it admits (the six MLX exports and their revisions), what it deliberately doesn't do (no engine, no inference), and that #1514 and #1508 build on it? Then it can leave draft.

[jschmied]

A suggestion for this PR, if it helps. The XPU backend for LFM3 (#1666) needs to load the text decoder by these names. weights.zig already has every one, self_attn.gate_proj included, but only inside schema().

Commit 865e50022 on jschmied/TensorFold:lfm3-names sits directly on this PR's head 3b140b7ff. It moves the table into families/lfm3/names.zig:

• Namespace: bare model. or wrapped language_model.model.;
• Module, with each module's path;
• moduleName(), detect() and embedName().

weights.zig re-exports them, and schema() and validate() read the same table, so admission is unchanged. names.zig doesn't import the safetensors parser. That matters because a backend's core module already owns core/safetensors.zig, and Zig refuses a second module rooted at the same file. A new test in weights_test.zig checks every module name of both namespaces against the schema. test-lfm3-host passes.

Cherry-pick it if you agree. Otherwise the XPU PR will carry it on top of this one.

[BlivionIaG]

Adopted jschmied's names.zig commit (authorship kept) and rebased the admission change onto it; test-lfm3-host passes on this head.

[ashhart]

Thanks @BlivionIaG and @jschmied. The description is what we needed, and the shared names table is a good call. It's gating to land now, with main's ling_host line kept beside yours in build.zig (the rebase had doubled your lfm3_host line; we kept the one after flashnext_batch_host, where your commit put it).

[ashhart]

Thank you @BlivionIaG. This is merged into main with your commits under your own name, so you're listed as a TensorFold contributor. It ships in the next release, credited to you in the changelog and the release notes.

[ashhart]

Thanks @BlivionIaG and @jschmied! LFM3-1.7B and LFM3-VL-2B's official MLX checkpoints are admitted and checked in 1.2.1: https://github.com/ashhart/TensorFold/releases/tag/v1.2.1