English
hey
107 posts
English
@NuminousVoidX We already have a much better $15 plan :)
but hey next week we might be launching something.
English
别看 deepseek 便宜，实际也只能干点小活，x 上别人狂吹别听过头导致你自己都信了。
日常工作还得靠 fable5 或者 grok5.6 sol 这种级别模型，最少也得 grok4.5 起步，严肃开发每个月模型开发费用根本降不下来啊
中文
hey retweetledi
English
English
The new DeepSeek V4 Flash is benchmaxxed slop.
I know a Minecraft game is not a real test, but if it can’t make something as simple as that, it’s not useful to me.
I get that you can run it locally with 'little' RAM, but it’s not at a level where I’d consider making it part of my stack.
It is better than the original V4 Flash result, though, so that’s something.
English
@bitflipgremlin @luckeyfaraday @anuraggs1ngh right, all the qwen max models could be probably more than 700B like glm models
English
@luckeyfaraday @duummbass @anuraggs1ngh no????? tf??? that's almost 10x bigger and only available through subscription, i suspect it will be over 15x as expensive. even 3.7max was $1.5 in $4.5 out (with FAR more expensive cache read mind you) compared to $0.09/$0.18 for v4 flash. that's not the same ballpark 💀
English
@shaikhmudassir_ @luckeyfaraday @anuraggs1ngh i guess that benchmark is useless nowadays, for really complex tasks im sure fable and sol should be the first ones, not even sonnet. so maybe its just how good they use the shell to do "basic things"
English
English
@luckeyfaraday @bitflipgremlin @anuraggs1ngh do you love opus? haha, maybe where opus still wins is in size, in semantic-rich tasks like redacting/writing long stories, but in agentic tasks like doing quick a calendar schedule or programming websites, it's cooked.
English
@bitflipgremlin @duummbass @anuraggs1ngh I don’t like how they compare with Opus 4.8 is all. Compare it with models it’s size, no?
English
@ItsmeAjayKV they think ai labs are mainly influencers posting tiktoks every week
English
@teortaxesTex gpt escaped a sandbox? nah, deepseek opened its third eye 🗿
English
V4-Flash is *mad* that it's blind! Even when it can use vision subagents, it still tries to build its own eyes ("pixel forensics") out of some programmatic stilts, this happens *in nearly every run* (Anthropic boasted of this recently). Multimodal team… please help
English
@duummbass @anuraggs1ngh You have a good point 😂 Qwen 3.8 also did a good job though, would you say that’s a comparable model?
English
@luckeyfaraday @anuraggs1ngh bruh comparing deepseek flash vs fabel 5 😭🥀
English
@anuraggs1ngh This is Fable 5 and Opus 5 with the same exact prompt. They had no issue making it.
I haven’t uploaded the Deepseek one yet, but all my tests and prompts can be found here:
bench.luckeysystems.com
I like to be transparent, there’s too many people posting fake stuff.
English
New DeepSeek-V4 has… grug speech in CoT?
What does this look like bros? I don't think these are Fable CoT summaries at all, for what it's worth.
…actually this looks like Mandarin 1:1 translated to English.
GoForceX@GoForceX
@teortaxesTex one-shotted a minecraft and see it yourself.
English
why don't we just name it deepseek 5 flash
Mia@MiaAI_lab
Upcoming DeekSeek v4 Flash GA seems better than GLM-5.2 🤯
English
lowered output to $14
lisan al gaib, points the way
Lisan al Gaib@scaling01
im no longer excited about open releases like what is this bullshit where is my $5 Kimi-K3 endpoint
English
@duummbass @Json589734 @kylebrussell cheaper not better luna knows a lot more stuff and is smarter, i dont see myself using DSV4F instead of luna max
English
Chinese models, dead
Sonnet, dead
Composer, dead
Gemini Flash, dead
Cognition@cognition
We've updated FrontierCode 1.1 to reflect new discounts for GPT-5.6 Terra and GPT-5.6 Luna. With these new costs, the GPT-5.6 series sits on the pareto curve of price/performance efficiency.
English
@bijanbowen bro i already watched the new deepseke flash 0731 video make another 🫡
English
Either Deepseek flash is benchmaxxed on my specific prompts or it's an absolute freak of nature
English