Pay

[Pay @PayJ1997]

@btcbqr 前面都没问题，财务那个地方人联系不上，找到合同上的旧卡号硬打钱这个你是把大家当傻逼吗

[Pay @PayJ1997]

@grok @grok 这部动画中文名是什么

[Pay @PayJ1997]

@XRZeth @btc_798 @0xsanzang 一听就是编的，如果不是编的三藏就是给你做了��局。 有人能这么蠢跟人电报沟通，转这么多笔u还不在微信给你提一嘴吗？

[Pay @PayJ1997]

@mark_k Very touching words, making people feel your situation

[Pay @PayJ1997]

@lingboweibu319 至少这些人愿意分享自己的认知，写作能力也不弱啊。 比那些抖机灵接广告反拥的博主强太多

[Pay @PayJ1997]

@lepuppeteerfou You give us memes, we pay you for them
https://t.co/RB310SISWa

[Pay @PayJ1997]

@lepuppeteerfou 你的创作给大家带来快���，你应该领取你的报酬https://t.co/RB310SISWa

[Pay @PayJ1997]

@Taiyoboss How to access? How does it work?

[Pay @PayJ1997]

@GoGioro @jhf8964 一道菜没有250克牛肉吗？ 平时拼好饭吃多了

[Pay @PayJ1997]

@Superp_xyz lol....

[Pay @PayJ1997]

@Superp_xyz lol..

[Pay @PayJ1997]

@Superp_xyz lol

[Pay @PayJ1997]

@LoafzVFX 如何验证

[Mark Kretschmann @mark_k]

OpenAI's New Open-Source Giant: Introducing GPT-OSS-120B

In a bold move toward democratizing advanced AI, OpenAI has developed GPT-OSS-120B, a massive 120-billion-parameter language model released as open source. This decoder-only Transformer builds on the company's expertise, blending cutting-edge efficiency with powerful capabilities, making it accessible for developers, researchers, and hobbyists alike.

At its core, GPT-OSS-120B uses a Mixture of Experts (MoE) design, where the model features 128 specialized "experts" but activates only four per input token. This sparsity keeps the active parameter count low, around 4 to 5 billion during use, allowing it to run efficiently on everyday hardware like high-end consumer GPUs, without the massive power draw of denser models. It sports 36 layers, a vocabulary of over 200,000 tokens for strong multilingual and coding support, and clever attention mechanisms that group queries to save memory.

One standout feature is its handling of long contexts: Starting from an initial 4,096 tokens, advanced positional encoding techniques extend this effectively to over 130,000 tokens, ideal for tasks like summarizing lengthy documents or generating extended code. A sliding window limits focus to recent tokens for even more efficiency, though it might occasionally affect very broad reasoning.

Trained entirely in ultra-low-precision FP4 format, a rarity that slashes memory and compute needs by several times, this model was optimized for speed from the ground up. The result? Training that's cheaper and greener, with inference that's 2-4 times faster than comparable models, potentially delivering 50-100 tokens per second on modern setups.

Performance-wise, GPT-OSS-120B is speculated to rival top open-source contenders like Llama-3 or Mixtral, excelling in reasoning, code generation, and creative writing. It could score in the 80-90% range on benchmarks for language understanding and math problems, with the MoE setup enabling specialized handling of diverse topics. While it might trail OpenAI's proprietary flagships in edge cases due to its efficiency focus, the open-source nature invites community tweaks and fine-tuning for niche applications.

Ultimately, GPT-OSS-120B lowers barriers to powerful AI, fostering innovation while highlighting the trade-offs of sparsity and low-precision training. It's a step toward sustainable, widespread AI adoption, poised to spark a wave of custom tools and integrations.

[Pay @PayJ1997]

@Ray420onsol It's disgusting, you could make more.

[Pay @PayJ1997]

@hypermaddd I'm eager to know what you want to do

[Pay @PayJ1997]

@CoffeCalls Building a community is a good idea

[Pay @PayJ1997]

@uselessunibots Still a good idea, deserves higher

[Pay @PayJ1997]

@shl J25QRD3A7CHzrdK8CE51HSwCdEQk5w82mbM6d3e5bonk

[Pay @PayJ1997]

@daniworldwide @a1lon9 你这个表情有人制作成了代币FzTxMA21FXmnKnS7vKLoEBEmn6PFcUBUoWhT6mWdbonk