Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Gemini 3.5 Flash

blog.google

Post reply on HN
Print view View original

681–690 of 692 posts

Re: Gemini 3.5 Flash

#681

Post by data-ottawa »

wwizo wrote: post #665
data-ottawa wrote: post #648

Anyone using this yet? I’m finding it very bad at instruction following vs 3.1. It calls tools it is told shouldn’t, and it loves calling tools. There’s a pretty strong bias towards its training vs system prompt instructions. Google’s release notes say to reduce unnecessary tool calls by reducing thinking, but that feels like it should be orthogonal to me. It definitely has improved a few logic things, like in data v…

Same. Feels very goal oriented. Requires multiple attempts to deter course and means to achieve it. On tool use. Gave it interactive design assignment on Antigravity 2. Failed miserably until I asked to use playwright for testing. And boy did it go with it. Tested hell out of visuals, nailed the solution. On following instruction. Asked Gemini Flash 3.5 to summarize YouTube video (google io developer keynote), a task…

What thinking level were you using?

In my testing, the minimal thinking mode hallucinated 2/3 times, which is pretty scary. The other modes weren’t as bad. I don’t have comprehensive data though.

Re: Gemini 3.5 Flash

#682

Post by LarsDu88 »

DCKing wrote: post #675
LarsDu88 wrote: post #674

Earlier quoted context omitted.

TurboQuant. They can fit more in less now

TurboQuant is a runtime optimization for a model's KV cache and doesn't allow for reduction in model size.

TurboQuant reduces the runtime memory needed for the model's KV cache.

This reduces both the memory bandwidth needed for inference (at the cost of slightly increasing the amount of compute needed), and the amount of VRAM used overall, meaning more VRAM can be allocated for more weights on the same hardware.

You were replying to a comment estimating model params from hardware. I am saying the param count could be higher for the same hardware.

Re: Gemini 3.5 Flash

#683

Post by Doxon »

smnscu wrote: post #530
easygenes wrote: post #466

For those who would like to know the total and active parameter count of this model: even though Google doesn't disclose the model technicals, we can infer them within relatively tight margins based on what we do know. We know they serve the model on TPU 8i, which we have plenty of hard specs for (so we know the key constraints: total memory and bandwidth and compute flops). We can also set a ceiling on the compute c…

Nice post! You piqued my curiosity, so after a bit of research it turns out that, with techniques like MTP/MLA/CSA, it's quite probable that these models are much more efficient (and maybe bigger? tho 400B sounds about right) than a simple RAM breakdown would suggest. MTP - https://blog.google/innovation-and-ai/technology/developers-... MLA - https://machinelearningmastery.com/a-gentle-introduction-to-... CSA - https…

These techniques are used by DeepSeek, and work well with the commodity (NVIDIA) GPU's they use. Google designs their entire AI stack from the custom silicon up. So they have different optimization approaches. (Though Gemma does use MTP)

Re: Gemini 3.5 Flash

#684

Post by balnaphone »

OhMeadhbh wrote: post #641
sagarpatil wrote: post #533

Earlier quoted context omitted.

Frontpage, Dreamviewer, flash, photoshop lol. We are old.

and Pagemill and Sitemill. At Bell Canada we had a very early web dev team in '94-'95. At one point pagemill came out and we could hire mostly non technical designers to build web pages. At the time it seemed like magic. We didn't need to have someone who grokked vi standing next to a designer all the time. But the HTML pagemill spat out was horrid. It always added a space to the end of link text and never closed lis…

Re: Sandhill. This literally sounds like BNR humour circa '94.

Re: Gemini 3.5 Flash

#685

Post by SXX »

segmondy wrote: post #667
SXX wrote: post #662

Earlier quoted context omitted.

What Kimi exactly? What version and quant?

K2.6, Qx/Q4, it's huge and mostly runs off CPU/system ram. So slow

I'd say its as good as Gemimi 3.1 Pro for this specific test.

Re: Gemini 3.5 Flash

#686

Post by -mlv »

gpugreg wrote: post #342
ai_fry_ur_brain wrote: post #220

Earlier quoted context omitted.

You dont understand the costs involved to run inference at scale Please go run some numbers.The hardware needed to Run Deepseek v4 flash at 20 tps for a single session is nowhere close to what is required to run it at 50tps for 5,000 concurrent sessions. Imagine what it takes to be profitible when running at 150 tps for 30cents per 1mm. You make less than 1k per month and the hardware required to run that cost 10k a…

> Please go run some numbers. - DeepSeek serves DeepSeek V4 Pro at 27 tps: https://openrouter.ai/deepseek/deepseek-v4-pro - At 27 tps per user, a B300 GPUS will give you around 800 tokens per second (serving 30 users): https://developer-blogs.nvidia.com/wp-content/uploads/2026/0... - That's 800 * 60 * 60 generated tokens per hour, at a cost of $0.87 per 1M tokens, or $2.50 per hour. - For input and output tokens, the…

$3.40 is for spot instances from what I'm seeing and you'll probably gonna need more than 1 for V4 Pro.

Re: Gemini 3.5 Flash

#687

Post by nothingfalsy »

BLAH BLAH BLAH. I don't trust anything the a company say how good their product is. try it yourself and see if its actual any good.

flash is barely good its okay but really shit on anything that matters

flash lite is absolute garbage. super stupidly retarded. I am going to die from high blood pressure on how stupid it is

Re: Gemini 3.5 Flash

#688

Post by NitiX »

WhitneyLand wrote: post #197
GodelNumbering wrote: post #84

Per million input/output tokens: Gemini 2.5 flash: $0.30/$2.50 Gemini 3.0 flash preview: $0.50/$3.00 Gemini 3.5 flash: $1.50/$9.00 Interesting pricing direction. I don't think we have ever seen a 3x price increase for in the immediate next same-sized model (and lol @ 3 only ever getting a preview). 3.5 flash costs similar to Gemini 2.5 pro which was $1.25/$10

Their rationale might be that it’s size and intelligence are growing relative to the market. Fwiw it’s beating Claude Sonnet in most benchmarking (benchmaxxing?), yet they’ve priced it almost half off on a per token basis. Question is are you going to persuade anyone with this argument? Are there many devs at Google who legit prefer Gemini over Claude and Codex? Would love to hear about that.

[flagged]

Re: Gemini 3.5 Flash

#689

Post by MASNeo »

esafak wrote: post #366
MASNeo wrote: post #128

Well, available for Gemini means these days that half the time they are “Receiving a lot of requests right now.” and so sorry they couldn’t complete the task. Luckily the model supports long time horizons because that’s what’s needed. /me likes Gemini a lot just wishing Google would add the compute!

Are you on a paid plan?

I was, yes.

Re: Gemini 3.5 Flash

#690

Post by SavioMak »

WarmWash wrote: post #632
DCKing wrote: post #583

Earlier quoted context omitted.

If two things hold up - 1) this is actually a 2-300B parameter model and 2) this is actually competitive with frontier OpenAI and Anthropic models (and not just benchmaxing), the implications are pretty big. It would mean you could run "frontier level" performance in one box at home. 300B models at least fit in a single maxed out Mac Studio or a small stack of DGX Sparks or AMD Strix Halo boxes. For comparison, DeepS…

Opus is estimated to be around 4T parameters, and 5.5 around 9T. [1] And while 3.5 at least qualifies to be in the same neighborhood, which is stunning if these numbers are all true, it may be that closing that last ~10% difference needs 50x more parameters. [1] https://arxiv.org/pdf/2604.24827

Note that this paper is vibe-coded and overestimating due to incorrect analysis, though "the core idea behind the paper is largely sound".

https://x.com/justanotherlaw/status/2050399317782155726 https://www.lesswrong.com/posts/veFMEzDDyWaer2Sms/sanity-che...

Post reply on HN