tiyuvta/Hy3-NVFP4

Hy3 NVFP4 — Memra × Tiyuvta

Weight-only NVFP4 safetensors quantization of Tencent Hy3, generated directly from the official BF16 checkpoint with NVIDIA Model Optimizer 0.46.0.

Generated with Memra, a native Rust and CUDA inference engine, by Tiyuvta (inference.tiyuvta.ai).

Artifact availability is not hosted-model availability. Tiyuvta does not currently publish a hosted Hy3 model id. This repository is an open model artifact, not a customer endpoint.

Provenance

Source	tencent/Hy3@a960ebc3da325ba167f069f76c41eb62c9280d22
Source format	BF16 safetensors, 99 shards
Quantizer	NVIDIA ModelOpt 0.46.0 @ 43fd41a58d52c4e6e5dec1d1ff5989ecc737ae1a
Recipe	Group-16, weight-only W4A16; fused gate/up pairs share max(gate_scale_2, up_scale_2)
Memra mint	ac1357fff29f23526729d832b249519c7c6d3dce
Memra qualified runtime	bfd5a2d88ef653576e2ea5bfc8711408eb2d1cac
Memra support declaration	bfd5a2d88ef653576e2ea5bfc8711408eb2d1cac (NativeQualified)
Minted artifact manifest	d63f8c3da9ab144d42dbfc1136d05294e26d5b7b7a40b114bf4b301359f4092c

The tokenizer, chat template, generation config, and license come from the pinned source. The official Tencent FP8 checkpoint is a separately pinned control, not the source for this mint.

Precision split

Surface	Stored precision
Routed-expert weights, trunk layers 1–79	NVFP4, 45,504 tensors
Routed-expert weights, MTP layer 80	NVFP4, all 576 tensors
Expert activations	BF16 (W4A16); not FP4- or INT8-quantized
Dense layer 0	BF16
Attention Q/K/V/O and QK norms	BF16
Router and correction bias	BF16/F32 source precision
Shared MLP	BF16
Embeddings, output head, norms, non-expert MTP tensors	BF16/F32 source precision
KV cache	Runtime-selected; not quantized by this artifact

There is no BF16-MTP variant or fallback in this release. The complete artifact contains 46,080 NVFP4 expert weights, 139,298 stored tensors normalized to 47,138 logical tensors, and an indexed tensor payload of 180,826,481,152 bytes.

Integrity locks

File/surface	SHA-256
config.json	3cb16aa29d0046ffddd2f8a4866e4c7511e4018c6fced8dd913d1a788d787af9
hf_quant_config.json	38e5689cd6847427cc28c26c3cd3ca30568822bf311f479f11d21cf8ab632d2e
model.safetensors.index.json	0f22f6fc51ac7e39b7510a77c77098c4fd7c722e9e6cfdb9782247c37f1b6afd
Mint logical tensor census	566db2975edac5cd1a86061ec6943988ef695cc8ae8c6cda050ad0d354ae2600
NativeQualified inspect census	1a1f7330e62153d24b17f8083f08ef1c314d90287ecf210f0b351b233f07a138
Minted 108-file manifest	d63f8c3da9ab144d42dbfc1136d05294e26d5b7b7a40b114bf4b301359f4092c

The repository includes SHA256SUMS for uploaded-byte verification. The minted-manifest hash above binds the quantizer output before the Hugging Face model card replaces the upstream README.md; the upstream README is retained as README_UPSTREAM.md.

Validation

The exact artifact was qualified with Memra on four NVIDIA RTX PRO 6000 Blackwell Server Edition cards. Memra's model-family-independent automatic loader compiled the canonical ModelPlan, bound the exact tensor census, and selected whole-expert parallelism for the four-card artifact. No HY3 layer list or architecture-specific parallel recipe was supplied.

• Same-artifact ModelOpt oracle (vLLM pipeline placement): matching argmax, 20/20 top-logit overlap, cosine similarity 0.9995285974311506, RMSE 0.05849238475772628, mean absolute error 0.046138540558894096, and maximum absolute error 0.30402064323425293.
• Automatic capacity planning selected PP-2 at the legal layer-41 cut, then EP-3 and EP-4 as the expert root fit with reserve. On four cards it estimated 60.05 GB on the root and 40.26 GB on each peer before the 6 GiB runtime reserve.
• Greedy MTP K=1–8: target-identical at every K, with nonzero accepted drafts at every K. The sampled Memra server also engaged MTP explicitly (9/22 accepted in the qualification request).
• Vendor-default sampling was tested with no sampling fields; Memra resolved Tencent's temperature 0.9 and top-p 1.0 defaults.
• Native tool-call parsing, tool-result turns, reasoning_effort high/none, concurrent requests, cache reuse, admission, client-disconnect rollback, peer integrity, and clean rollback to the plain server all passed.

MTP is qualified but remains an explicit serving choice. MEMRA_PP_WAVE also remains off for this artifact: its experimental schedule improved throughput internally, but failed the required serial-vs-wave logit-identity gate for the mixed BF16/NVFP4 numeric program. No internal aggregate throughput result is promoted into a public performance claim.

Memra tuning is not finished. This artifact is NativeQualified, not NativeTuned; the remaining generic 2–4 card kernel, TP+EP, routing, and MTP work is tracked in Memra issue #67.

Run with Memra

CUDA_VISIBLE_DEVICES=0,1,2,3 \
MEMRA_PARALLEL=auto \
MEMRA_PARALLEL_DEVICES=0,1,2,3 \
MEMRA_SERVE_SPEC=0 \
MEMRA_SERVE_BATCH=1 \
MEMRA_MODELS="hy3=/path/to/Hy3-NVFP4" \
memra-server

Qualified runtime binaries:

Binary	SHA-256
memra-server	46f65e0699a9ed9871d8bb648361ddf8ddfb96080ee6aad4a77a58c9a3325d85
run-spec	4c856bf67bab710dc8f3fd3493885fe732dc4185168e3986fd887733ea62b244
run-safetensors	99d637845d2b2cd01b1c8935e30de198db375f6a1bab3d542bd60cfabb618557

The published sampling defaults follow Tencent's configuration: temperature 0.9 and top-p 1.0. Clients can override them per request.

Files

• 99 indexed safetensors shards;
• config.json, hf_quant_config.json, model.safetensors.index.json;
• tokenizer, tokenizer config, chat template, and generation config;
• Apache-2.0 license and retained upstream READMEs;
• SHA256SUMS covering the uploaded files (excluding itself).

License

Hy3 is licensed under Apache-2.0. This artifact carries the pinned upstream license.

Credits

• Tencent Hy Team for Hy3.
• NVIDIA for ModelOpt and the NVFP4 format.
• Memra for the native loader, tensor contract, and serving engine.
• Tiyuvta for artifact generation and RTX Blackwell qualification.