NHacker Next
  • new
  • past
  • show
  • ask
  • show
  • jobs
  • submit
▲Tencent Hy4 Flash (hy.tencent.ai)
chriddyp 2 days ago [-]
This one scored very well on our data analytics benchmarks [1].

Beat out GPT-6 Luna on the low end of cheap + accurate. Only cost $0.13 to answer 43 data analytics questions (compared to ~$20 on the high end of expensive frontier models).

Once the open the weights, Hy4 Flash and regular Hy4 will be the only two open models on the frontier in the exam I've been running.

[1] https://plotly.com/blog/hy4-flash-plotly-data-analytics-benc...

ddxv 2 days ago [-]
Cheaper pricing than Deepseek 1017 and Luna and supposedly 'faster' (though I guess in practice a lot of that comes down to where it's running.

Lately I let myself drift into using Terra and some other mildly 'expensive' models, but really I think the sweet spot for me is cheap and fast.

jitl 2 days ago [-]
I think DeepSeek 1017 max is better than Terra at creativity and autonomy. I really like that model.
DeathArrow 1 days ago [-]
I'm pretty amazed by Deepseek 1017.

While testing it with Pi, for $10 I've built a pretty complex app, I built the UI, a CI/CD, ran manual testing in a browser, did load testing, fined tuned it to get a decent RPS with low memory consumption.

Probably Opus would have done it faster, but not for $10.

Next step in testing is using Opus for planning and review and Deepseek for coding.

nolok 1 days ago [-]
> supposedly 'faster' (though I guess in practice a lot of that comes down to where it's running.

In LLM terms "faster" means one or two (or both) things usually, either it's inference is faster, or it gets to its answer in less tokens overall.

alfiedotwtf 1 days ago [-]
FWIW faster means less active parameters but not too small, all while still being Q3 or above...

A3B is probably too small, and A23B is probably too big, but Qwen 3.8's Flash Next at A6B seems about juuuuust right

throwa356262 23 hours ago [-]
Qwen 4 flash will be really interesting. 3.8FN was able to run on really weak hardware, it just needed a tiny bit more intelligence to compete with the big boys.
yasindce1998 13 hours ago [-]
I am running this on OpenCode (Right now it is free in opencode) and getting around 250t/s. It's doing a very decent job.
tonfreed 2 days ago [-]
I use the expensive ones at work where I don't have to pay for them, but DeepSeek and GLM flash models with the occasional k3.1 pro code review works really well for me
DeathArrow 1 days ago [-]
It's cheaper than Deepseek if you look at the oficial Deepseek pricing. But there are lots of providers that are cheaper than that.
thesurlydev 2 days ago [-]
I've been using this model via openrouter for the last couple of days in place of deepseek flash with good results.
tkslucas 2 days ago [-]
Lovely, it's always great to see new model drops
syntaxing 2 days ago [-]
Wow 224B with 64B anagrams. It might be possible to run this well on two 128GB machine. I’ve been exclusively using deepseek flash 1017 as my “smart” model through open router. Really curious how this compares
anon373839 2 days ago [-]
Hmm. The ngrams can be offloaded to SSD, leaving 160B that need to reside in memory. I'm wondering if this can fit and perform well at ~3 bpw on a single 128GB machine. At least, can it beat Qwen Flash Next at 4 bpw?
ycui7 2 days ago [-]
this can run on a single DGX Spark if nvidia make a 4-bit QAD local quant.
wsxiaoys 2 days ago [-]
it's doing ok on my multimodal/3D task: https://sketchagent.dev/s/Ss6HgYEq reasoning looks solid and properly accept visual feedback.

More results: https://sketchagent.dev/#/a/hy4-flash

jauntywundrkind 2 days ago [-]
Benchmarks are competitive with glm-5.3 flash, and man! That price! Ridiculously cheap. Engrams really rocking it, eh?

224b params. 320b for glm-5.3-flash. 180b for qwen3.8 flash next. 552b for DeepSeek flash v4.1 but that has way stronger performance, for 4x+ the price.

Also, they do mention at the end that yes they are going to open the weights here.

Kuyawa 2 days ago [-]
I love the smell of fresh competition, it brings prices down and features up, that's the essence of free markets, the real reason behind the "War on AI"
edtechdev 1 days ago [-]
I thought I could use it for some translation work, but it made many more errors than DeepSeek 4.1 Flash 1017
trilogic 1 days ago [-]
Interesting claims, very low cost according to the tables. On my way to test it.
1 days ago [-]
throwa356262 24 hours ago [-]
You would assume the company behind tiktok has enough compute to serve this but their inference speed and reliability has not been the best so far:

Eden reports 40tps for hy4: https://www.edenai.co/models/hy4

And error rate is not insignificant according to Openrouter: https://openrouter.ai/tencent/hy4

ketzu 2 days ago [-]
Curious to test it, but Hy4 via openrouter was such a nightmare in reliability for me when testing it for basic error correction. A niche usecase, but unlike most models I had so many broken responses there.

Maybe Hy4 flash will at least not also be expensive while breaking on me.

2 days ago [-]
DeathArrow 1 days ago [-]
There is also this cheap one from Baidu worth checking out: https://ernie.baidu.com/release/ernie-5-5-flash-preview/
DeathArrow 1 days ago [-]
I don't think it beats Deepseek v4.1 Flash 1017 at cost per task. Deepseek is really great and there are some really cheap providers if you look for them.

Even a $10 Open Code plan will get you far because it will hit cache a lot if you use a matching harness.

Rodmine 2 days ago [-]
The smaller size seems like a win. While nothing beats Deepseek 4.1 flash on price/performance, I hope the smaller size means that 3rd party providers can provide much faster speeds (DS 4.1 flash is often available around 200-400 tps in Openrouter), and hence there will be actual competition.
vcryan 2 days ago [-]
I love it. It's a bit slow. Hopefully other providers will pick it up.
paramseast 1 days ago [-]
[dead]
TheJCDenton 2 days ago [-]
[dead]
scottchiefbaker 2 days ago [-]
[dead]
italiansolider 2 days ago [-]
[dead]
Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact
Rendered at 20:10:41 GMT+0000 (Coordinated Universal Time) with Vercel.