# `.basalt` file format changelog The version is the `u32` after the magic in the file's header. The engine reads one version; the [model file](model-file.md) page describes the current one. ## Version 2 The current version. The engine refuses version 1 files and names [tools/migrations/basalt_v1_to_v2.py](../tools/migrations/README.md), which writes a version 2 file from a version 1 file without a download or any re-quantizing. - **Contains:** `meta`, `tokenizer/*`, `dense`, `dense.index`, `experts`, `experts.layout`, `gguf`, `mtp/*`, `hc8`, `dspark/gguf`, `vision/*`. - **Added `hc8`, now required:** the int8 copies of the hyper-connection matrices the decode reads, the bits the engine used to compute at every start. A start reads them as they are and skips reading the BF16 matrices to quantize them (about half a second). - **Removed `profile`:** the expert profile decides which experts start in VRAM. It is placement policy that improves independently of the weights, so it moved to the engine (`data/profiles`), where a better profile needs no new file. - **`experts.layout` v7, no version change:** a layer whose gate and up rows differ in type (Unsloth's UD-IQ1_S: IQ2_S gate and IQ2_XXS up on one layer) gives the up type as a sixth column under an `# experts v7` header. The converter writes v7 only for such a model; every other pack keeps v5 (GGUF) or v6 (NVFP4). An engine from before v7 refuses a v7 pack and asks for an update. - **Tensor type 45, `NVFP4_PLANES`, no version change:** NVFP4 as a code plane and 128 x 4-tiled UE4M3 scales (the TMA layout; [model file](model-file.md#nvfp4_planes-tensor-type-45)), the FP32 scale a separate `.scale`. The Qwen3.8-27B pack's MLP and head use it; the engine refuses that pack's older ggml-block form. - **Tensor type 44, `F8_E4M3`, no version change:** one E4M3 code per value with the matrix's FP32 scale as a separate `.scale` tensor (F32 [1]), as NVIDIA's FP8 checkpoints store them. The Qwen3.8-27B pack's projections use it. An engine from before it refuses such a tensor by name ("the dense kernels read ..."). - **Added `dspark/gguf`, optional, no version change:** a DeepSeek pack's DSpark drafter. The engine ignores sections it does not name, so an engine from before it runs such a pack plain, and a pack without it runs plain on any engine. - **Trimmed `meta`:** no `sampling` (the server reads the GGUF metadata's `general.sampling.*` keys), no `profile` or `hc8` records and no converter revision, provenance that changed whenever a section was replaced. ## Version 1 - **Contains:** `meta` (with `sampling`, `profile` and `hc8` records and the converter's git revision), `tokenizer/*`, `dense`, `dense.index`, `experts`, `experts.layout`, `gguf`, `mtp/*`, `profile`, optionally `hc8`, `vision/*`. - The first single-file format: everything the engine, the server and the vision sidecar load. Without `hc8` the engine quantized the hyper-connection matrices at every start.