Hy-Memory-30B-A3B — Radiance .rad artifacts
Radiance-native inference containers derived from Tencent/Hy-Memory-30B-A3B. These are .rad files for the Radiance inference engine, not Transformers/safetensors, GGUF, or vLLM checkpoints. The source model is licensed under Apache-2.0; its license is included here. Radiance software notices are provided separately.
Artifacts
| File | Size | Contents |
|---|---|---|
hy-memory-30b-a3b-bf16.rad |
60,136,292,352 bytes (60.14 GB; 56.00 GiB) | BF16 reference container |
hy-memory-30b-a3b-w4-gptq-full-1m.rad |
16,598,839,296 bytes (16.60 GB; 15.45 GiB) | Full-GPTQ W4: attention, layer-0 dense MLP, and routed experts are calibrated; shared MLP, router, norms, embeddings, and output head remain BF16 |
The full-GPTQ artifact uses asymmetric U4 for attention and the layer-0 dense MLP, and signed I4 routed experts, with group size 128. Its calibration Grams, recipe, and provenance are in the companion calibration dataset repository.
Serving
Use a Radiance build with HYV3 support and the shipped hyv3_bf16 plugin. The BF16 artifact was smoke-tested at TP2. The W4 artifact was tested at TP1 and TP2; an example TP1 launch is:
radiance --model ./hy-memory-30b-a3b-w4-gptq-full-1m.rad --tp 1
The W4 artifact completed a synthetic 10,015-token prefill and 2,000-token decode on one AMD Radeon AI PRO R9700 at 5,620.6 prefill tokens/s and 89.94 decode tokens/s. This is a throughput measurement, not a quality result.
Quantization fidelity (held-out, teacher-forced)
Evaluation used Radiance's tokenizer and 192 origin-disjoint documents (code, math, prose, and tools), scoring 66,086 next-token positions. KL is full-vocabulary KL(BF16 || candidate); top-1 is agreement with the BF16 model.
| Variant | Mean KL vs BF16 | Top-1 agreement | Candidate PPL (BF16 PPL 4.1169) |
|---|---|---|---|
| BF16 reference | 0 | 100% | 4.1169 |
| RTN W4 | 0.18494 | 89.22% | 4.3287 |
| Expert-only GPTQ W4 | 0.16320 | 89.86% | 4.2683 |
| Full-GPTQ W4 | 0.10711 | 92.50% | 4.1826 |
These are corpus-specific fidelity measurements, not a guarantee of broad task superiority. Full GPTQ improved aggregate held-out likelihood versus RTN, but regressed on the held-out tools-source traces.
Task and tool-use results
Greedy exact-task results: GSM8K 183/200 (91.5%), HumanEval 131/164 (79.9%), and MBPP sanitized 173/257 (67.3%) for full-GPTQ W4. Paired intervals versus RTN included zero; no broad task-quality win is established.
The native OpenAI tools API was evaluated on a frozen, safe 113-case executable BFCL v3 subset. Calls were parsed by Radiance, executed by a local deterministic allowlist, and results were replayed through the API as tool messages. All cases executed without API errors.
| Variant | Exact call matches |
|---|---|
| BF16 | 104/113 (92.0%) |
| Full-GPTQ W4 | 105/113 (92.9%) |
| RTN W4 | 103/113 (91.2%) |
Full GPTQ versus RTN had two GPTQ-only wins and no RTN-only wins (exact paired McNemar p=0.50). Final-answer prose was not independently correctness-graded; this small subset does not establish an overall quality winner.
License and attribution
The model weights derive from Tencent's Hy-Memory-30B-A3B, licensed under Apache-2.0; see LICENSE. Radiance-related software notices are in LICENSE-RADIANCE and NOTICE-RADIANCE. This repository distributes model artifacts; it does not grant any additional rights beyond the applicable upstream licenses.