Modelpublic

Baekpica/Motif-3-GGUF

sourceHugging Facemitupdated 28d agoView on Hugging Face

0likes33downloads

Model Card

Motif 3 — Q8_0 GGUF

Support my work

I work on making large language models practical on hardware they were never really designed to fit on — through mixed quantization, inference optimization, custom kernels, and serving experiments.

While much of the development happens on local hardware, calibration, profiling, and large-scale validation often require expensive on-demand GPUs.

Contributions help pay for that compute, storage, and testing infrastructure, so I can keep experimenting and publishing the results openly.

Full-topology Q8_0 reference conversion of `Motif-Technologies/Motif-3`. This is an independent conversion, not an official Motif Technologies release.

The conversion retains all 53 transformer layers, including the first two dense layers and all 51 sparse layers; all 384 routed experts per sparse layer (top-8 routing); the shared expert; Grouped Differential Latent Attention (GDLA); Expert-Specific PolyNorm; modified mHC; and the one-layer MTP head. Nothing is pruned, merged, expert-dropped, or layer-dropped.

Artifact

Variant	Split	Exact size	Purpose
Q8_0 reference	11 shards	334,810,734,464 bytes (311.817 GiB)	imatrix calibration and full-model correctness reference

Start with Motif-3-Q8_0-00001-of-00011.gguf; split-aware runtimes discover the remaining shards automatically. Exact hashes are in Q8_0-SHA256SUMS.

The weight files are fixed at Hub revision 5c266c95bf8c8d822d50e5e1cce9d108eaadb2af. Later model-card commits do not change shard bytes or hashes.

bash

hf download Baekpica/Motif-3-GGUF \
  --revision 5c266c95bf8c8d822d50e5e1cce9d108eaadb2af \
  --include 'Motif-3-Q8_0-*.gguf' \
  --include Q8_0-SHA256SUMS \
  --local-dir ./Motif-3-Q8_0

This reference follows the control-path protection policy used by the mixed release. Its 2,287 GGUF tensors comprise:

Tensor type	Count	Role
`Q8_0`	696	embeddings, LM head, GDLA/dense/shared/MTP projections, all routed expert gate/up/down matrices
`BF16`	318	small mHC projection matrices
`F32`	1,273	routers, norms, PolyNorm coefficients/biases, and mHC controls/scalars

The protected BF16/F32 tensors are intentional. “Q80 reference” means every large quantizable weight matrix, including every routed expert matrix, is Q80; decision-sensitive controls remain at their locked higher precision.

Provenance


Source model	`Motif-Technologies/Motif-3`
Exact source revision	`ccceb1a5fd7b5eb32e47841216b3caf5666c07bc`
Source parameters	314,841,775,750
Source tensors	2,236
GGUF tensors	2,287 (51 fused gate/up tensors are losslessly separated)
GGUF architecture	`motif3`
Native context metadata	262,144 tokens
Fixed Q8-weight revision	`5c266c95bf8c8d822d50e5e1cce9d108eaadb2af`
Official implementation oracle	`MotifTechnologies/vllm@4cd9eb4129883565e69d508038d783d59ee01867`
Conversion base	`ggml-org/llama.cpp@1d2869c6e54d5003f3927a79efbca0fefa034a6d`
Native runtime	`Baekpica/ds4:feature/motif-3-model-loader@d878ea1`
Public reproduction	`Baekpica/motif-3-mixed-ds4`
Private Spark handoff	Expensive state is preserved in `hf://buckets/Baekpica/motif-3-spark-handoff`

Only the official final Motif-3 checkpoint above was used. Motif-3-Beta was not used as a source, calibration input, implementation oracle, or fallback.

Validation status

The completed 11-shard artifact passed a strict source-map verification:

—all 2,287 expected GGUF tensors are present exactly once;
—every shape, type, byte length, and split index/count matches the locked map;
—all 53 layers are present, including routed gate/up/down tensors for sparse layers 2–52 and the full MTP block;
—general.source.revision is the pinned official revision above;
—the aggregate verified tensor payload is 334,801,074,712 bytes;
—no structural-template marker remains in the release files.

A separate source-row sanity pass dequantized 57 representative rows spanning the embedding, LM head, GDLA, dense/shared paths, routed gate/up/down matrices at sparse layers 2/26/52, protected controls, and MTP. It also samples both halves of the fused source gate/up tensors across experts 0/173/383. F32 and BF16 protected rows were source-exact; Q8_0 cosine was at least 0.9999740124, with maximum sampled relative RMSE 0.007201264.

The machine-readable report and complete per-shard SHA-256 list are included as Q8_0-VERIFY.json, Q8_0-SAMPLE-VERIFY.json, and Q8_0-SHA256SUMS.

The 262,144-token value is source architecture metadata. This Q8 repository does not by itself claim completed single-DGX-Spark 256K serving. That release gate belongs to the native motif3 ds4 runtime and requires measured resident weights, latent KV, workspace/server memory, prefill, decode, and semantic checks on GB10.

Runtime compatibility

Motif 3 is not a Llama-family graph. A runtime must implement its 384E sigmoid router and route normalization, shared expert, Expert-Specific PolyNorm, GDLA/differential heads and output gate, modified mHC, interleaved SWA/full attention with YaRN, latent KV semantics, and MTP. Do not assume that a stock GGUF runtime recognizes motif3 merely because it can parse the container.

The H200 development branch is publicly available as `Baekpica/ds4:feature/motif-3-model-loader` at exact implementation commit d878ea1a1d67bc0f0bd60e20e75b4a011aa2d8d9. The private Spark handoff also preserves an offline source snapshot and commit metadata.

Intended use in the mixed-quant pipeline

This artifact is produced first and frozen as the full-model Q80 reference. The next stage collects an activation importance matrix from the pinned calibration corpus, after which routed gate/up weights are quantized to `IQ2XXS with that imatrix and routed down weights to Q2_K. The resulting MQ87–88 artifact is published separately at [Baekpica/Motif-3-Mixed-Quant-GGUF`](https://huggingface.co/Baekpica/Motif-3-Mixed-Quant-GGUF).

License and attribution

The source model identifies its license as MIT. See the official `Motif-Technologies/Motif-3` model card for the architecture, intended-use, evaluation, citation, and license context.