Tencent Hy3 weights have just gone up on HuggingFace:
It is a 295B-A21B model, so good size for 2x Spark configs.
The benchmark results are compelling:
71.7 on TerminalBench-2.1 and 28 on DeepSWE are standout results for this size class. In many of the results it sits somewhere between GLM-5.1 and GLM-5.2.
No NVFP4 releases yet, so keep your eye out.
Vitals:
| Property | Value |
|---|---|
| Architecture | Mixture-of-Experts (MoE) |
| Total Parameters | 295B |
| Activated Parameters | 21B |
| MTP Layer Parameters | 3.8B |
| Number of Layers (excluding MTP layer) | 80 |
| Number of MTP Layers | 1 |
| Attention Heads | 64 (GQA, 8 KV heads, head dim 128) |
| Hidden Size | 4096 |
| Intermediate Size | 13312 |
| Context Length | 256K |
| Vocabulary Size | 120832 |
| Number of Experts | 192 experts, top-8 activated |
| Supported Precisions | BF16 |
Just seen this. A W4A16/Int4/NVFP4 quant should easily fit 2 Sparks and may indeed be more capable than DSV4 Flash if the real-world performance is as strong as the benchmarks promise.
21b active parm’s is nice. I often notice the low active parm models suffer more from lacking nuance and forgetting things.
Now that Deep Seek Flash is running well, lol, let’s go for it!
I tried for fonzies Hy3 preview highly quantized on Mac - its was unbelievably nerfed and bland - cannot be explained by quantizing, not incoherent - its was all good in a little scope I tested, just incredibly RLHF’ed and corporate-styled. Sorta like Gemini or Nemotron model: “I cannot provide this advise - consult professionals”. Will see how it will differ in released version.
Has anyone tried this version?
It’s GQA and the base preview model’s been out for a while, so I think it shouldn’t be hard to run it. I don’t have any disk space left so I’ll wait until someone posts a benchmark before trying it myself : )
Would it be possible to somehow fit this into a single spark at some quant?
Something like a 2 bit or 3 bit quant gguf should fit. But even if it fits the performance would be pretty misrable at 21B active tokens. Probably in the low teens for decode and a few hundred for prefill. I think you’ll be better off with Step 3.7 Flash for one spark.
I am going to adapt PrismaQuant and AURA to use GGUF’s format catalog and see if I can create a version that runs on a single spark. No promises but fingers crossed. It won’t leverage native nvfp4 but that will probably be OK
A standard vLLM PrismaAURA targeting quality + full context on 2x Sparks would also be a great contribution! I think there’s a bit of room over straight NVFP4 to experiment (if KV cache is relatively efficient).
Alas, I only have a single spark, so testing would be difficult.
