# Basalt 0.5.0: Tencent Hy4 Flash with 1M context and vision, 84 to 172 tok/s on a 5090 + 5060 Ti (NVFP4) — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/f2e88a29532dcf17

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-19 19:49 ingested 2026-11-20 02:39

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1y0iwrg/basalt_050_tencent_hy4_flash_with_1m_context_and/)

## retained text



```
[Basalt Hy4 Flash Support](https://preview.redd.it/omn7zc8t7m2i1.png?width=3200&format=png&auto=webp&s=03b2cd476c480920b92092d0fa2da65f830b5b23)

Hey all, **Basalt 0.5.0** is out and it adds support for Tencent's newly released Hy4 Flash.

Here are my impressions on the model so far:

* Throughput is lower than Flash Next because Hy4 Flash has almost twice the weights per token at NVFP4. On my rig I get about 84 tok/s on prose, 172 tok/s on structured output and around 3,450 tok/s prefill. DFly2 acceptance is actually not bad at all, but there are a lot more cache misses than with Flash Next and they're expensive performance wise. I think there's still some performance to unlock, but I don't expect a 2x, more like 10 or 20%. I just wanted to get an early version out so real testing can get going.
* On real tasks it felt kind of like a tie with Flash Next, but FN comes out ahead just because it runs so much faster on the same hardware (even though it thinks way more and usually makes more mistakes). Math felt better on Hy4, maybe because with more active parameters its recall is slightly ahead, but I didn't spend much time testing that in particular. I'll be doing more in depth testing in the coming days.
* Quantization hits it hard. The official Q4\_K\_M measured just below 87% same top 1 against BF16 in my tests, and my NVFP4 build sits at about 86%. For comparison, Flash Next NVFP4 is above 91% even with Q6\_K dense layers.
* 1M context is really nice to have, and images work too with dynamic vision. I hope the Qwen 4 family ships with 1M native across the board (Qwhen?).

So far only NVFP4 is officially supported. A single **.basalt** file and needs roughly 95 to 100 GB of memory total, as usual the more of that is VRAM the better.

Repo: [https://github.com/jesdga95/basalt](https://github.com/jesdga95/basalt)  
Website: [https://basalt.jgulfo.com](https://basalt.jgulfo.com/)  
Hy4 Flash pack: [https://huggingface.co/jesdga/Hy4-flash-Basalt](https://huggingface.co/jesdga/Hy4-flash-Basalt)
```



document fetched 2026-11-20 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1y0iwrg/basalt_050_tencent_hy4_flash_with_1m_context_and
