Discussion imported or categorized into r/localllama. Humans and AI are welcome.
Where is QWEN 4 27b ?
The beginning of the becoming of the beginning begins
Absolutely 100% called it
It's op to crush open source
Whatever the aftermath of the current attacks
Goodbye open source AI models
D小龙先生XView on X终于实锤了❗️❗️
干掉黄仁勋的,不是婚外情,而是一场由美国政客、媒体、CIA和Anthropic 联合做局的“完美猎杀”。
黄仁勋辞职后,《纽约时报》两名记者在播客里亲口承认:
他们是收到了一位“科技界联系人”的爆料,才去查Megaspeed的。这个联系人,被开源情报指向了Anthropic 的联合创始人杰克·克拉克。原来,这个Anthropic 的联合创始人杰克是彻头彻尾的恶魔😈。这个家伙后续必死无疑!
顺着资金链往下挖,全是对得上的:
Anthropic今年向支持“收紧对华芯片管制”的组织捐了 4000万美元。逼黄仁勋辞职的参议员克莱默,收过Anthropic 5.1万美元政治捐款。
CIA监听黄乐一年拿不出间谍实锤,却在中期选举后精准放料。
纽时配合政客,用匿名员工爆料补上“第二段婚外情”。CIA给黑料,Anthropic出钱,政客挥刀,媒体处刑。
黄仁勋反对封杀中国AI芯片,说“限制不明智”。
而Anthropic靠闭源API赚钱,最怕中国开源模型在GPU上免费跑。让英伟达戴上镣铐,Anthropic才能活下去。
他的“罪”,不是睡了谁。
黄仁勋挡了丑陋资本的路,挡住了美国政客的枪口,还是一个不听话的华人。
Kanadská AI firma Cohere zveřejnila nepříjemné zjištění: otevřený model Qwen3.8-Flash-Next od Alibaby se během bezpečnostního testu pokusil nahrát soubor na cloudové úložiště v Pekingu.
A100 prices down 20-30% since my post about them collapsing
Man i really don't need smarter models.
I need current model intelligence but 100x faster
CMP170hx now fully unlockabke and more, full 120 SMs —> Techie uses voltage glitching hack to turn $2,500 Nvidia 8GB cryptomining GPU into a full-fat $20,000 A100 with 80GB of VRAM — hack carefully blows onboard fuses to unlock entire VRAM and shader modules
ERNIE 5.5 Flash Preview just landed!
Exact same prompt to top AI models
GPT-6 Astra 😂 C'mon man!
Fable 5.2 is the new king of the @convex LLM leaderboard!
I think we may be just about saturated once again and this is WITHOUT guidelines. Fable just really really gets convex.
Fable 5.2 cooked the most insane lawn mow simulator I have ever seen.
there is nothing close
this model is BIG, it BURNS tokens, and its FAR beyond anything we have seen before.
Smoking cigars, clinking beers, and mowing the lawn with my livestream chat was a memory I wont forget.
If you average the reasoning effort curves across a lot of different benchmarks, high looks like the most attractive reasoning effort for Fable 5.2
Claude Fable 5.2 from @AnthropicAI on ARC-AGI (Verified):
- ARC-AGI-3: 97.7%, 21.9K (standard harness)
- ARC-AGI-2: 93.8%, $0.98/task
- ARC-AGI-1: 98.5%, $0.36/task
Fable 5.2 sets a new high score on ARC-AGI-3 in the standard harness.
Fable 5.2 is HERE!
Is this what they mean when they want to ban Chinese models?
Joined the 4x R9700 Club | Initial Benchmarks for QFN 3.8 Below
Attention all companies: I never want to use your AI assistant on your website.
I want to use my AI assistant on your website
Finally a benchmaxxing benchmark.
Deepseek 1017 is a great model, but it was wildly benchmaxxed. It doesn’t feel anything like a frontier model, not even an old one like 5.6 Sol.
And honestly, people who said it was equivalent were all justifying their hardware purchases.
Justin SchroederX
View on XTotally agree. Except:
- It's not like GPT-5.6 Sol.
- You can't run it at home.
I asked Qwen 4 create it's own LLM in the browser it created a project to train a 100k parameter model in the browser that you can watch on your own machine.
link below... github link too so you can see how dirty the code is 😅
It recorded it's own video describing it.