Tencent Hy3 (July 2026 production release, Apache 2.0) — 295B total /
21B active MoE, 80 layers, 192 routed experts top-8, including the MTP
draft layer for speculative decoding — quantized by
PrismaQuant from the BF16 source
to a ~190 GB compressed-tensors checkpoint (5.28 bits/param over
quantizable weights) sized to serve on two NVIDIA DGX Sparks under
vLLM pipeline parallelism with no custom kernels and no forked runtime.
Sibling artifact: the single-Spark 2.8-bit GGUF is at
rdtand/Hy3-295B-A21B-PrismaQuant-2.8bit-gguf-vllm.
allocation/layer_config.json.model.layers.80 ships verbatim BF16 and loads into
vLLM's HYV3 MTP path (--speculative-config '{"method":"mtp",...}').No quality claims are made for this artifact, and the author has one Spark: the full 2-node topology has not been run. What was validated:
serving/tp2_shard_audit.py)
verifies every quantized tensor's packed dims, NVFP4 group-16 boundaries,
and scale shapes divide cleanly at world size 2.Community measurements (2×Spark throughput, KL/perplexity vs BF16, task suites) are welcome and will be linked here.
See serving/README.md for the two-node pp2 launch (recommended) and the
tp2 alternative.
Quantized with PrismaQuant. Source: tencent/Hy3 (BF16). Allocation and
Pareto data in allocation/. Contact: robert.tand@icloud.com
model-00006-of-00038.safetensors 5.4 GBmodel-00002-of-00038.safetensors 5.4 GBmodel-00036-of-00038.safetensors 5.4 GBmodel-00031-of-00038.safetensors 5.4 GBmodel-00004-of-00038.safetensors 5.4 GBmodel-00007-of-00038.safetensors 5.4 GBmodel-00037-of-00038.safetensors 5.4 GBmodel-00005-of-00038.safetensors 5.4 GBmodel-00023-of-00038.safetensors 5.4 GBmodel-00017-of-00038.safetensors 5.4 GBmodel-00028-of-00038.safetensors 5.4 GBmodel-00034-of-00038.safetensors 5.4 GBmodel-00001-of-00038.safetensors 5.4 GBmodel-00008-of-00038.safetensors 5.4 GBmodel-00027-of-00038.safetensors 5.4 GBmodel-00003-of-00038.safetensors 5.4 GBmodel-00033-of-00038.safetensors 5.4 GBmodel-00014-of-00038.safetensors 5.4 GBmodel-00018-of-00038.safetensors 5.4 GBmodel-00024-of-00038.safetensors 5.4 GBmodel-00029-of-00038.safetensors 5.4 GBmodel-00016-of-00038.safetensors 5.4 GBmodel-00022-of-00038.safetensors 5.4 GBmodel-00035-of-00038.safetensors 5.4 GBmodel-00026-of-00038.safetensors 5.4 GBmodel-00015-of-00038.safetensors 5.4 GBmodel-00021-of-00038.safetensors 5.4 GBmodel-00009-of-00038.safetensors 5.4 GBmodel-00019-of-00038.safetensors 5.4 GBmodel-00012-of-00038.safetensors 5.4 GBmodel-00032-of-00038.safetensors 5.4 GBmodel-00030-of-00038.safetensors 5.4 GBmodel-00010-of-00038.safetensors 5.4 GBmodel-00020-of-00038.safetensors 5.4 GBmodel-00011-of-00038.safetensors 5.4 GBmodel-00025-of-00038.safetensors 5.4 GBmodel-00013-of-00038.safetensors 5.4 GBmodel-00038-of-00038.safetensors 5.1 GBHave the files? Submit a torrent.
Submit a torrentHave the complete files and permission to share them? Replace the example folder path. The script checks the files, makes a torrent and submits its magnet and checksums. Keep your torrent client seeding. Get your community key from your settings, and read the script before running it.
curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo rdtand/Hy3-295B-A21B-PrismaQuant-5.3bit-2xSpark-vllm ./model-foldertencent/Hy3