Tencent Hy4 Flash

(hy.tencent.ai)

127 points | by lyra1337 13 hours ago ago

22 comments

  • chriddyp 10 hours 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 9 hours 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 5 hours ago

      I think DeepSeek 1017 max is better than Terra at creativity and autonomy. I really like that model.

      • DeathArrow an hour 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 2 hours 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.

    • DeathArrow 2 hours 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.

    • tonfreed 8 hours 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

  • trilogic 29 minutes ago

    Interesting claims, very low cost according to the tables. On my way to test it.

  • thesurlydev 10 hours ago

    I've been using this model via openrouter for the last couple of days in place of deepseek flash with good results.

  • edtechdev an hour ago

    I thought I could use it for some translation work, but it made many more errors than DeepSeek 4.1 Flash 1017

  • syntaxing 7 hours 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 4 hours 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 3 hours ago

      this can run on a single DGX Spark if nvidia make a 4-bit QAD local quant.

  • DeathArrow an hour ago

    There is also this cheap one from Baidu worth checking out: https://ernie.baidu.com/release/ernie-5-5-flash-preview/

  • jauntywundrkind 12 hours 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.

  • DeathArrow an hour 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.

  • ketzu 4 hours 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.

  • Kuyawa 10 hours 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"

  • wsxiaoys 10 hours 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

  • tkslucas 10 hours ago

    Lovely, it's always great to see new model drops

  • Rodmine 10 hours 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 10 hours ago

    I love it. It's a bit slow. Hopefully other providers will pick it up.