Русский
English Indonesia Türkçe Deutsch Español Português Pусский Français Italiano Nederlands Polski العربية ไทย Tiếng Việt 繁體中文 简体中文 日本語 한국어
lm_zheng's profile picture. Member of technical staff @xAI | Prev: Ph.D. @UCBerkeley, Co-founder @lmsysorg

Lianmin Zheng

@lm_zheng

Member of technical staff @xAI | Prev: Ph.D. @UCBerkeley, Co-founder @lmsysorg

491Посты 14Kчитателей 637в читаемых
Вам может понравиться
Lianmin Zheng сделал(а) репост
haoailab's profile picture.
Hao AI Lab
@haoailab
19 нояб.г.

(1/N) 🚀 We converted a high quality Wan2.2-MoE into an autoregressive model. Preview checkpoint: huggingface.co/FastVideo/Caus… - First autoregressive version of Wan2.2-A14B MoE - I2V compatible - 8-step distilled - Potential backbone for streaming generation and world modeling Try…

haoailab's tweet card. FastVideo/CausalWan2.2-I2V-A14B-Preview-Diffusers · Hugging Face
FastVideo/CausalWan2.2-I2V-A14B-Preview-Diffusers · Hugging Face
Источник: huggingface.co
8
53
266
198
35К
Lianmin Zheng сделал(а) репост
lmsysorg's profile picture.
LMSYS Org
@lmsysorg
19 нояб.г.

🚀 Introducing Miles — an enterprise-facing RL framework for large-scale MoE training & production, forked from slime. Slime is a lightweight, customizable RL framework that already powers real post-training pipelines and large MoE runs. Miles builds on slime but focuses on new…

lmsysorg's tweet image. 🚀 Introducing Miles — an enterprise-facing RL framework for large-scale MoE training & production, forked from slime. Slime is a lightweight, customizable RL framework that already powers real post-training pipelines and large MoE runs. Miles builds on slime but focuses on new…
lmsysorg's tweet image. 🚀 Introducing Miles — an enterprise-facing RL framework for large-scale MoE training & production, forked from slime. Slime is a lightweight, customizable RL framework that already powers real post-training pipelines and large MoE runs. Miles builds on slime but focuses on new…
lmsysorg's tweet image. 🚀 Introducing Miles — an enterprise-facing RL framework for large-scale MoE training & production, forked from slime. Slime is a lightweight, customizable RL framework that already powers real post-training pipelines and large MoE runs. Miles builds on slime but focuses on new…
7
21
157
108
50К
Lianmin Zheng сделал(а) репост
GenAI_is_real's profile picture.
Chayenne Zhao
 @GenAI_is_real
19 нояб.г.

We have implemented unified FP8 RL. FP8 for both training and rollout effectively eliminates train–inference inconsistency caused by quantization error, improving both the speed and stability of RL training.

GenAI_is_real's tweet image. We have implemented unified FP8 RL. FP8 for both training and rollout effectively eliminates train–inference inconsistency caused by quantization error, improving both the speed and stability of RL training.
GenAI_is_real's tweet image. We have implemented unified FP8 RL. FP8 for both training and rollout effectively eliminates train–inference inconsistency caused by quantization error, improving both the speed and stability of RL training.
GenAI_is_real's tweet image. We have implemented unified FP8 RL. FP8 for both training and rollout effectively eliminates train–inference inconsistency caused by quantization error, improving both the speed and stability of RL training.
GenAI_is_real's tweet image. We have implemented unified FP8 RL. FP8 for both training and rollout effectively eliminates train–inference inconsistency caused by quantization error, improving both the speed and stability of RL training.
2
11
109
45
8К
Lianmin Zheng сделал(а) репост
GenAI_is_real's profile picture.
Chayenne Zhao
 @GenAI_is_real
17 нояб.г.

We conducted additional experiments to compare speculative training with frozen MTP layers and obtained solid results. Further experiments are being conducted on larger-scale (300B+) MOEs, looking forward!

GenAI_is_real's tweet image. We conducted additional experiments to compare speculative training with frozen MTP layers and obtained solid results. Further experiments are being conducted on larger-scale (300B+) MOEs, looking forward!
GenAI_is_real's tweet image. We conducted additional experiments to compare speculative training with frozen MTP layers and obtained solid results. Further experiments are being conducted on larger-scale (300B+) MOEs, looking forward!
GenAI_is_real's profile picture. Learning and Leading RL at SGLang @lmsysorg BS @Tsinghua_Uni ｜ PhD @UCLA
Chayenne Zhao
 @GenAI_is_real
14 нояб.г.

We introduce speculative decoding into the RL sampling process, achieving a significant improvement in sampling speed under appropriate batch sizes. Compared to freezing the draft model, the accepted length maintain at a high level, generating long-term stable positive gains.

GenAI_is_real's tweet image. We introduce speculative decoding into the RL sampling process, achieving a significant improvement in sampling speed under appropriate batch sizes. Compared to freezing the draft model, the accepted length maintain at a high level, generating long-term stable positive gains.
GenAI_is_real's tweet image. We introduce speculative decoding into the RL sampling process, achieving a significant improvement in sampling speed under appropriate batch sizes. Compared to freezing the draft model, the accepted length maintain at a high level, generating long-term stable positive gains.
GenAI_is_real's tweet image. We introduce speculative decoding into the RL sampling process, achieving a significant improvement in sampling speed under appropriate batch sizes. Compared to freezing the draft model, the accepted length maintain at a high level, generating long-term stable positive gains.
GenAI_is_real's tweet image. We introduce speculative decoding into the RL sampling process, achieving a significant improvement in sampling speed under appropriate batch sizes. Compared to freezing the draft model, the accepted length maintain at a high level, generating long-term stable positive gains.
4
28
222
106
23К
0
12
61
24
10К
Lianmin Zheng сделал(а) репост
elonmusk's profile picture.
Elon Musk
@elonmusk
17 нояб.г.

Grok 4.1 just released. You should notice a significant increase in speed and quality.

xai's profile picture.
xAI
 @xai
17 нояб.г.

Introducing Grok 4.1, a frontier model that sets a new standard for conversational intelligence, emotional understanding, and real-world helpfulness. Grok 4.1 is available for free on grok.com, grok.x.com and our mobile apps. x.ai/news/grok-4-1

xai's tweet card. Grok 4.1 is now available to all users on grok.com, 𝕏, and the iOS and Android apps. It is rolling out immediately in Auto mode and can be selected explicitly as “Grok 4.1” in the model picker.
Grok 4.1 | xAI
Источник: x.ai
1К
3К
12К
1К
32.1М
11К
11К
61К
3К
29.1М
lm_zheng's profile picture.
Lianmin Zheng
@lm_zheng
17 нояб.г.

Grok 4.1 is out! Amazing improvements.

xai's profile picture.
xAI
 @xai
17 нояб.г.

Introducing Grok 4.1, a frontier model that sets a new standard for conversational intelligence, emotional understanding, and real-world helpfulness. Grok 4.1 is available for free on grok.com, grok.x.com and our mobile apps. x.ai/news/grok-4-1

xai's tweet card. Grok 4.1 is now available to all users on grok.com, 𝕏, and the iOS and Android apps. It is rolling out immediately in Auto mode and can be selected explicitly as “Grok 4.1” in the model picker.
Grok 4.1 | xAI
Источник: x.ai
1К
3К
12К
1К
32.1М
13
15
307
7
42К
lm_zheng's profile picture.
Lianmin Zheng
@lm_zheng
14 нояб.г.

Numeric debugging is extremely challenging given the huge number of kernels in today’s training and inference systems. You can’t get a single kernel wrong. This is a great engineering achievement and will hopefully make RL training and debugging much easier.

lmsysorg's profile picture. Large Model Systems Organization: Join our Slack: https://t.co/mSPNyKTLTS We developed SGLang https://t.co/jEqIJcGwGA, Chatbot Arena (now @lmarena_ai), and Vicuna!
LMSYS Org
@lmsysorg
14 нояб.г.

💥 We've achieved perfect training-inference alignment for SGLang & FSDP in slime! (Flash Attn 3, DeepGEMM, etc.) The result? A strict KL divergence of 0. But here's the twist: We spent a month trying to find a baseline that crashes from mismatch... and couldn't. 🤷‍♂️ We haven't…

lmsysorg's tweet image. 💥 We've achieved perfect training-inference alignment for SGLang & FSDP in slime! (Flash Attn 3, DeepGEMM, etc.) The result? A strict KL divergence of 0. But here's the twist: We spent a month trying to find a baseline that crashes from mismatch... and couldn't. 🤷‍♂️ We haven't…
lmsysorg's tweet image. 💥 We've achieved perfect training-inference alignment for SGLang & FSDP in slime! (Flash Attn 3, DeepGEMM, etc.) The result? A strict KL divergence of 0. But here's the twist: We spent a month trying to find a baseline that crashes from mismatch... and couldn't. 🤷‍♂️ We haven't…
lmsysorg's tweet image. 💥 We've achieved perfect training-inference alignment for SGLang & FSDP in slime! (Flash Attn 3, DeepGEMM, etc.) The result? A strict KL divergence of 0. But here's the twist: We spent a month trying to find a baseline that crashes from mismatch... and couldn't. 🤷‍♂️ We haven't…
lmsysorg's tweet image. 💥 We've achieved perfect training-inference alignment for SGLang & FSDP in slime! (Flash Attn 3, DeepGEMM, etc.) The result? A strict KL divergence of 0. But here's the twist: We spent a month trying to find a baseline that crashes from mismatch... and couldn't. 🤷‍♂️ We haven't…
11
23
240
92
77К
5
16
194
45
37К
Lianmin Zheng сделал(а) репост
lmsysorg's profile picture.
LMSYS Org
@lmsysorg
14 нояб.г.

💥 We've achieved perfect training-inference alignment for SGLang & FSDP in slime! (Flash Attn 3, DeepGEMM, etc.) The result? A strict KL divergence of 0. But here's the twist: We spent a month trying to find a baseline that crashes from mismatch... and couldn't. 🤷‍♂️ We haven't…

lmsysorg's tweet image. 💥 We've achieved perfect training-inference alignment for SGLang & FSDP in slime! (Flash Attn 3, DeepGEMM, etc.) The result? A strict KL divergence of 0. But here's the twist: We spent a month trying to find a baseline that crashes from mismatch... and couldn't. 🤷‍♂️ We haven't…
lmsysorg's tweet image. 💥 We've achieved perfect training-inference alignment for SGLang & FSDP in slime! (Flash Attn 3, DeepGEMM, etc.) The result? A strict KL divergence of 0. But here's the twist: We spent a month trying to find a baseline that crashes from mismatch... and couldn't. 🤷‍♂️ We haven't…
lmsysorg's tweet image. 💥 We've achieved perfect training-inference alignment for SGLang & FSDP in slime! (Flash Attn 3, DeepGEMM, etc.) The result? A strict KL divergence of 0. But here's the twist: We spent a month trying to find a baseline that crashes from mismatch... and couldn't. 🤷‍♂️ We haven't…
lmsysorg's tweet image. 💥 We've achieved perfect training-inference alignment for SGLang & FSDP in slime! (Flash Attn 3, DeepGEMM, etc.) The result? A strict KL divergence of 0. But here's the twist: We spent a month trying to find a baseline that crashes from mismatch... and couldn't. 🤷‍♂️ We haven't…
11
23
240
92
77К
lm_zheng's profile picture.
Lianmin Zheng
@lm_zheng
10 нояб.г.

Great progress!

casper_hansen_'s profile picture. NLP Scientist | AutoAWQ Creator | Open-Source Contributor
Casper Hansen
@casper_hansen_
10 нояб.г.

insane blackwell progress in v0.5.5 by the sglang team. with new optimizations, it's stable like hopper and the performance is great even for multimodal models 181 tokens/s on Qwen3-VL-30B-A3B-Thinking on 1x B200:

casper_hansen_'s tweet image. insane blackwell progress in v0.5.5 by the sglang team. with new optimizations, it's stable like hopper and the performance is great even for multimodal models 181 tokens/s on Qwen3-VL-30B-A3B-Thinking on 1x B200:
1
9
72
11
18К
2
0
69
5
7К
lm_zheng's profile picture.
Lianmin Zheng
@lm_zheng
7 нояб.г.

Hao has been pioneering efficient architecture research for many years. Always eager to see the innovations from him and his group!

haozhangml's profile picture. Asst. Prof. @HDSIUCSD and @ucsd_cse running @haoailab. Cofounder and runs @lmsysorg. 20% with @Snowflake
Hao Zhang
@haozhangml
7 нояб.г.

Excited to partner with SGLang: FastVideo + SGLang = the future open ecosystem for diffusion. 🥳🫡 ----------- A few extra cents: Since I started faculty at UCSD, our lab has been investing diffusion for video and text , and in both algorithms and systems. - Text-side, we…

haozhangml's tweet card. TL;DR: LLMs have been traditionally regarded as sequential decoders, decoding one token after another. In this blog, we show pretrained LLMs can be easily taught to operate as efficient parallel...
Consistency Large Language Models: A Family of Efficient Parallel Decoders
Источник: hao-ai-lab.github.io
3
8
103
27
26К
0
3
84
11
13К
lm_zheng's profile picture.
Lianmin Zheng
@lm_zheng
7 нояб.г.

Great progress!

slime_framework's profile picture. The LLM post-training framework for RL Scaling. https://t.co/4ILpx8hfKN
slime
 @slime_framework
7 нояб.г.

Ant AQ-Team @AQ_MedAI @TheInclusionAI and SGLang RL Team @sgl_project just helped land Kimi-K2-Instruct RL on slime — fully wired up and running on 256× H20 141GB 🚀 Huge shout-out to @yngao016, @menlzy, @Yonah_x from AQ Team and @Ji_Li_233, @Yefei_RL from the SGLang RL Team for…

slime_framework's tweet card. As the title says, I run it for 40 steps, and the raw_rewards is shown in the following figure: This work is done in collaboration with @GeLee-Q @yefei12 and @yzlnew
Add kimi-k2-instruct running script by Gao016 · Pull Request #694 · THUDM/slime
Источник: github.com
0
9
76
24
48К
0
1
24
2
4К
Lianmin Zheng сделал(а) репост
lmsysorg's profile picture.
LMSYS Org
@lmsysorg
7 нояб.г.

🚀 Introducing SGLang Diffusion — bringing SGLang’s high-performance serving to diffusion models. ⚡️ Up to 5.9× faster inference 🧩 Supports major open-source models: Wan, Hunyuan, Qwen-Image, Qwen-Image-Edit, Flux 🧰 Easy to use via OpenAI-compatible API, CLI & Python API…

lmsysorg's tweet image. 🚀 Introducing SGLang Diffusion — bringing SGLang’s high-performance serving to diffusion models. ⚡️ Up to 5.9× faster inference 🧩 Supports major open-source models: Wan, Hunyuan, Qwen-Image, Qwen-Image-Edit, Flux 🧰 Easy to use via OpenAI-compatible API, CLI & Python API…
lmsysorg's tweet image. 🚀 Introducing SGLang Diffusion — bringing SGLang’s high-performance serving to diffusion models. ⚡️ Up to 5.9× faster inference 🧩 Supports major open-source models: Wan, Hunyuan, Qwen-Image, Qwen-Image-Edit, Flux 🧰 Easy to use via OpenAI-compatible API, CLI & Python API…
lmsysorg's tweet image. 🚀 Introducing SGLang Diffusion — bringing SGLang’s high-performance serving to diffusion models. ⚡️ Up to 5.9× faster inference 🧩 Supports major open-source models: Wan, Hunyuan, Qwen-Image, Qwen-Image-Edit, Flux 🧰 Easy to use via OpenAI-compatible API, CLI & Python API…
lmsysorg's tweet image. 🚀 Introducing SGLang Diffusion — bringing SGLang’s high-performance serving to diffusion models. ⚡️ Up to 5.9× faster inference 🧩 Supports major open-source models: Wan, Hunyuan, Qwen-Image, Qwen-Image-Edit, Flux 🧰 Easy to use via OpenAI-compatible API, CLI & Python API…
15
66
317
156
99К
lm_zheng's profile picture.
Lianmin Zheng
@lm_zheng
7 нояб.г.

Future models will be multi modal in multi modal out, potentially combining auto regressive and diffusion architectures. SGLang project takes the first step towards building a unified inference stack for all.

lmsysorg's profile picture. Large Model Systems Organization: Join our Slack: https://t.co/mSPNyKTLTS We developed SGLang https://t.co/jEqIJcGwGA, Chatbot Arena (now @lmarena_ai), and Vicuna!
LMSYS Org
@lmsysorg
7 нояб.г.

🚀 Introducing SGLang Diffusion — bringing SGLang’s high-performance serving to diffusion models. ⚡️ Up to 5.9× faster inference 🧩 Supports major open-source models: Wan, Hunyuan, Qwen-Image, Qwen-Image-Edit, Flux 🧰 Easy to use via OpenAI-compatible API, CLI & Python API…

lmsysorg's tweet image. 🚀 Introducing SGLang Diffusion — bringing SGLang’s high-performance serving to diffusion models. ⚡️ Up to 5.9× faster inference 🧩 Supports major open-source models: Wan, Hunyuan, Qwen-Image, Qwen-Image-Edit, Flux 🧰 Easy to use via OpenAI-compatible API, CLI & Python API…
lmsysorg's tweet image. 🚀 Introducing SGLang Diffusion — bringing SGLang’s high-performance serving to diffusion models. ⚡️ Up to 5.9× faster inference 🧩 Supports major open-source models: Wan, Hunyuan, Qwen-Image, Qwen-Image-Edit, Flux 🧰 Easy to use via OpenAI-compatible API, CLI & Python API…
lmsysorg's tweet image. 🚀 Introducing SGLang Diffusion — bringing SGLang’s high-performance serving to diffusion models. ⚡️ Up to 5.9× faster inference 🧩 Supports major open-source models: Wan, Hunyuan, Qwen-Image, Qwen-Image-Edit, Flux 🧰 Easy to use via OpenAI-compatible API, CLI & Python API…
lmsysorg's tweet image. 🚀 Introducing SGLang Diffusion — bringing SGLang’s high-performance serving to diffusion models. ⚡️ Up to 5.9× faster inference 🧩 Supports major open-source models: Wan, Hunyuan, Qwen-Image, Qwen-Image-Edit, Flux 🧰 Easy to use via OpenAI-compatible API, CLI & Python API…
15
66
317
156
99К
6
23
213
45
18К
Lianmin Zheng сделал(а) репост
lmsysorg's profile picture.
LMSYS Org
@lmsysorg
6 нояб.г.

Day-0 support for Kimi K2 Thinking on SGLang ⚡ The new open-source thinking-agent model pushes reasoning, coding, and multi-step tool use to new heights. Proud to collaborate with @Kimi_Moonshot to make it run seamlessly: python -m sglang.launch_server \ --model-path…

lmsysorg's tweet image. Day-0 support for Kimi K2 Thinking on SGLang ⚡ The new open-source thinking-agent model pushes reasoning, coding, and multi-step tool use to new heights. Proud to collaborate with @Kimi_Moonshot to make it run seamlessly: python -m sglang.launch_server \ --model-path…
Kimi_Moonshot's profile picture. Built by Moonshot AI to empower everyone to be superhuman. ⚡️API: https://t.co/ggYlFf809H
Kimi.ai
@Kimi_Moonshot
6 нояб.г.

🚀 Hello, Kimi K2 Thinking! The Open-Source Thinking Agent Model is here. 🔹 SOTA on HLE (44.9%) and BrowseComp (60.2%) 🔹 Executes up to 200 – 300 sequential tool calls without human interference 🔹 Excels in reasoning, agentic search, and coding 🔹 256K context window Built…

Kimi_Moonshot's tweet image. 🚀 Hello, Kimi K2 Thinking! The Open-Source Thinking Agent Model is here. 🔹 SOTA on HLE (44.9%) and BrowseComp (60.2%) 🔹 Executes up to 200 – 300 sequential tool calls without human interference 🔹 Excels in reasoning, agentic search, and coding 🔹 256K context window Built…
584
2К
10К
4К
4.7М
1
6
27
0
5К
Lianmin Zheng сделал(а) репост
LiuYunlong63318's profile picture.
Yunlong Liu
@LiuYunlong63318
30 окт.г.

Awesome step to support native JAX inference from sglang! Now we don't need any torch layer to land code directly on TPU. @SingularMattrix

lm_zheng's profile picture. Member of technical staff @xAI | Prev: Ph.D. @UCBerkeley, Co-founder @lmsysorg
Lianmin Zheng
@lm_zheng
30 окт.г.

SGLang now has a pure Jax backend, and it runs natively on TPU!

2
5
159
23
21К
3
2
112
12
11К
lm_zheng's profile picture.
Lianmin Zheng
@lm_zheng
30 окт.г.

SGLang now has a pure Jax backend, and it runs natively on TPU!

lmsysorg's profile picture. Large Model Systems Organization: Join our Slack: https://t.co/mSPNyKTLTS We developed SGLang https://t.co/jEqIJcGwGA, Chatbot Arena (now @lmarena_ai), and Vicuna!
LMSYS Org
@lmsysorg
30 окт.г.

SGLang now runs natively on TPU with a new pure Jax backend! SGLang-Jax leverages SGLang's high-performance server architecture and uses Jax to compile the model's forward pass. By combining SGLang and Jax, it delivers fast, native TPU inference while maintaining support for…

lmsysorg's tweet image. SGLang now runs natively on TPU with a new pure Jax backend! SGLang-Jax leverages SGLang's high-performance server architecture and uses Jax to compile the model's forward pass. By combining SGLang and Jax, it delivers fast, native TPU inference while maintaining support for…
4
20
96
29
55К
2
5
159
23
21К
Lianmin Zheng сделал(а) репост
lmsysorg's profile picture.
LMSYS Org
@lmsysorg
28 окт.г.

Exciting to see Glyph open-sourced -- exploring a direction similar to DeepSeek-OCR, using visual-text compression to scale context windows! Instead of reading text token by token, Glyph lets models see the text, achieving 3–4× compression while preserving strong performance,…

Zai_org's profile picture. The AI lab behind GLM models, dedicated to inspiring the development of AGI to benefit humanity. https://t.co/b6zGxJvzzS
Z.ai
 @Zai_org
27 окт.г.

Glyph: Scaling Context Windows via Visual-Text Compression Paper: arxiv.org/abs/2510.17800 Weights: huggingface.co/zai-org/Glyph Repo: github.com/thu-coai/Glyph Glyph is a framework for scaling the context length through visual-text compression. It renders long textual sequences into…

Zai_org's tweet image. Glyph: Scaling Context Windows via Visual-Text Compression Paper: arxiv.org/abs/2510.17800 Weights: huggingface.co/zai-org/Glyph Repo: github.com/thu-coai/Glyph Glyph is a framework for scaling the context length through visual-text compression. It renders long textual sequences into…
38
94
686
279
87К
1
4
24
4
4К
Lianmin Zheng сделал(а) репост
jon_durbin's profile picture.
Jon Durbin
@jon_durbin
28 окт.г.

FYI, our SGLang fork is public here: github.com/chutesai/sglang (branch chutes) to boost your SGLang perf from 73% to 97% 👀 When I tested a few manually there were a few discrepancies where instead of generating a string it generated an array of strings, etc.; curious if switching…

jon_durbin's tweet card. SGLang is a fast serving framework for large language models and vision language models. - chutesai/sglang
GitHub - chutesai/sglang: SGLang is a fast serving framework for large language models and vision...
Источник: github.com
Kimi_Moonshot's profile picture. Built by Moonshot AI to empower everyone to be superhuman. ⚡️API: https://t.co/ggYlFf809H
Kimi.ai
@Kimi_Moonshot
28 окт.г.

Kimi K2vv updated! We've added case-by-case statistics for ToolCall-Trigger Similarity and ToolCall-Schema Accuracy. Feedback is welcome! github.com/MoonshotAI/K2-…

Kimi_Moonshot's tweet image. Kimi K2vv updated! We've added case-by-case statistics for ToolCall-Trigger Similarity and ToolCall-Schema Accuracy. Feedback is welcome! github.com/MoonshotAI/K2-…
10
32
398
90
93К
5
11
66
12
12К
Lianmin Zheng сделал(а) репост
lmsysorg's profile picture.
LMSYS Org
@lmsysorg
27 окт.г.

This huge contribution from @Baidu_Inc team enabled multi token prediction for Spare attention, achieving more than 2x decoding throughput improvements for the latest DeepSeek v3.2 models. The new architecture makes inference more interesting—we need to carefully handle the…

lmsysorg's tweet image. This huge contribution from @Baidu_Inc team enabled multi token prediction for Spare attention, achieving more than 2x decoding throughput improvements for the latest DeepSeek v3.2 models. The new architecture makes inference more interesting—we need to carefully handle the…
lmsysorg's tweet image. This huge contribution from @Baidu_Inc team enabled multi token prediction for Spare attention, achieving more than 2x decoding throughput improvements for the latest DeepSeek v3.2 models. The new architecture makes inference more interesting—we need to carefully handle the…
2
16
85
39
23К
YiMaTweets's profile picture. Chair Prof. in AI, HKU; Visiting Prof. of EECS, UCB For a Unified Theory of Intelligence: https://t.co/leZlkURb7j

Yi Ma

@YiMaTweets
iskyzh's profile picture. 这个人很菜，什么也不会 / 面对命中注定的失败

迟猫猫🐱

@iskyzh
karpathy's profile picture. Building @EurekaLabsAI. Previously Director of AI @ Tesla, founding team @ OpenAI, CS231n/PhD @ Stanford. I like to train large deep neural nets.

Andrej Karpathy

@karpathy
tqchenml's profile picture. AssistProf @CarnegieMellon. Distinguished Eng @NVIDIA. Creator of @XGBoostProject, @ApacheTVM. Member https://t.co/QYyfjQNp4p, @TheASF. Views are on my own

Tianqi Chen

@tqchenml
_akhaliq's profile picture. AI research paper tweets, ML @Gradio (acq. by @HuggingFace 🤗) dm for promo ,submit papers here: https://t.co/UzmYN5XOCi

AK

@_akhaliq
songhan_mit's profile picture.

Song Han

@songhan_mit
DrJimFan's profile picture. NVIDIA Director of Robotics & Distinguished Scientist. Co-Lead of GEAR lab. Solving Physical AGI, one motor at a time. Stanford Ph.D. OpenAI's 1st intern.

Jim Fan

@DrJimFan
BeidiChen's profile picture. Asst. Prof @CarnegieMellon, @amazon Scholar, Prev: Visiting Researcher @Meta, Postdoc @Stanford, Ph.D. @RiceUniversity, Large-Scale ML, a fan of Dota2.

Beidi Chen

@BeidiChen
cHHillee's profile picture. @thinkymachines Formerly @PyTorch "My learning style is Horace twitter threads" - @typedfemale

Horace He

@cHHillee
hardmaru's profile picture. Building @SakanaAILabs 🧠

hardmaru

@hardmaru
MarisaVeryMoe's profile picture. https://t.co/ATOeiQRiaT

Marisa "Long Middle Name Looks Cool" Kirisame

@MarisaVeryMoe
b1antaidaye's profile picture. Father of 2 | PhD @UCLAComSci | AssistProf @cemseKAUST | Compilers | Computer Arch | Sw/hw Co-designed Abstraction

Dr. Jian "dAIye" Weng

@b1antaidaye
gaocegege's profile picture. Co-founder and CEO @TensorChord, building postgres-based vector extension https://t.co/7WGvl1sR56 | Father of 1 cat | Married

Ce Gao

@gaocegege
tydsh's profile picture. ex-Research Scientist Director in Meta FAIR. Reasoning, Optimization and Understanding LLM. Novelist in spare time. PhD in @CMU_Robotics.

Yuandong Tian

@tydsh
Tim_Dettmers's profile picture. Creator of bitsandbytes.Research Scientist @allen_ai and incoming professor @CarnegieMellon. I blog about deep learning and PhD life at https://t.co/Y78KDJJFE7.

Tim Dettmers

@Tim_Dettmers
moyix's profile picture. Building offsec agents: https://t.co/G9EtnC2Gl3 PGP https://t.co/3WXr0RfRkv

Brendan Dolan-Gavitt

@moyix
yyw2000's profile picture. 🏳️‍⚧️ System Hacker.

vickieGPT

@yyw2000
soumithchintala's profile picture. Building new things @thinkymachines Cofounded and lead @PyTorch at Meta. Also dabble in robotics at NYU. AI is delicious when it is accessible and open-source.

Soumith Chintala

@soumithchintala
1SHL10's profile picture. 3rd-year PhD @PrincetonCS PL Group; PL/Systems; Prev @AWSCloud @Intel @Taichi_Lang @uwplse

¬¬Mike (Deyuan) He

@1SHL10
luisceze's profile picture. computer architect. marveled by biology. professor @uwcse. ceo @OctoAICloud. venture partner @madronaventures.

Luis Ceze

@luisceze
darrenlu666888's profile picture.

darrenlu

@darrenlu666888
lillian_mding's profile picture.

Lillian Ma

@lillian_mding
Nicholascelt's profile picture. AI Artist & Storyteller. Building the epic fantasy world of Nick Celt ⚔️. Videos, art, and adventure, all powered by AI. Follow the saga!

Nick Celt

@Nicholascelt
newjp114514's profile picture. esk . AI / @xAI @GoogleAI @SQLserver @Windows @bing @Apple /becomes Dr./

a天p

@newjp114514
weichen54370285's profile picture.

david.chen

@weichen54370285
kennetpostigo's profile picture. professional keyboard user

Kennet Postigo

@kennetpostigo
YiangJean93's profile picture.

li yian g

@YiangJean93
gham1lt0n's profile picture. Investing. Previously: @variantfund, Millennium

Geoff Hamilton

@gham1lt0n
KodaClement's profile picture. EE building Rockets 🚀 and big fan of anything Trains 🚆- Photographing Life - Opinions are my own

Clem

@KodaClement
1314LFP's profile picture. Blue Star Rover, the snooper in the corner

飞鹏

@1314LFP
Ayaan_Shariif's profile picture. engineer @XfiniteOfficial building @tensor_craft

Ayaan Sharif

@Ayaan_Shariif
xiangmitchell's profile picture.

Mitchell Xiang

@xiangmitchell
zoom_will's profile picture. Just a regular person from Earth 🌌🌌🌌

WK

@zoom_will
FawazAlQah39141's profile picture.

Fawaz AlQahtani

@FawazAlQah39141
tngc029's profile picture. TNG - Germany, UK, CH, ANZ. We're solving your hard IT problem

tngc029

@tngc029
Varbral911's profile picture.

Varbral

@Varbral911
DecadeZZZZ's profile picture.

DecadeZ

@DecadeZZZZ
Yassir1611's profile picture. on a mission to understand existence.

YA.AGI

@Yassir1611
qdeelta's profile picture.

QDelta

@qdeelta
NicoSchulzOrbs's profile picture. current teacher, future founder, creator and believer

Nico Schulz

@NicoSchulzOrbs
Elormchris4's profile picture. #Talk less,Do more! 🔱⚡🔌🔋₿₿₿ #Cooking https://t.co/eBoTQl9oYG The most amazing AI in the space! 👀👀👀 click https://t.co/eBoTQl9oYG and check out!

SⓂ️ARTⓂ️ONEY_RESEARCH ₿🧠🧠 Ikwe.ai @X

@Elormchris4
AliReza93582052's profile picture.

Ali Reza

@AliReza93582052
oaktownboy510's profile picture. #TSLA #Oakland #Raiders #Warriors #A’s

Sagar Patel

@oaktownboy510
DavidjHays's profile picture.

David Hays

@DavidjHays
joserivera234's profile picture. Tratando de aportar mi granito de arena para rescatar a la juventud contemporánea.

Jose Angel Rivera

@joserivera234
GecerIsa's profile picture.

Isa geçer

@GecerIsa
leonusa88's profile picture. Energy Business Consultant | Raised and Entrusted with Multi-Million Dollar Funds | Investment Advisor

Leon Ng

@leonusa88
ashleynguyeenx's profile picture. dive in to..my world

Ashley

@ashleynguyeenx
inderdeepsingh_'s profile picture. thesingh.eth | Software Engineer | Farmer | Building @Stablecoinxyz

Inderdeep Singh Sidhu

@inderdeepsingh_
billionaireops's profile picture. Fix your frequency.

Billionaire Protocol

@billionaireops
BearlyThear's profile picture. Catching sarcasm & passive-aggression AI misses 😏 Held-out samples + human scoring → beating frontier models 50–80% , data from real X posts. (more soon)

Bearly Thear

@BearlyThear
violazhouyi's profile picture. reporter @restofworld covering artificial intelligence and more. formerly @viceworldnews @scmpnews. 📧 viola@restofworld.org. signal: viola.572

Viola Zhou

@violazhouyi
DomOfTheWoods's profile picture.

Literally

@DomOfTheWoods
sabarish_sn's profile picture.

Sabarish

@sabarish_sn
anthonycxc's profile picture. Products & Creative Lead Web & EC, UI/UX, Branding, Visual #webflow #wordpress #figma Loves #tech #ai #food #coffee #wine #future

Anthony C

@anthonycxc
bonanzaeric's profile picture. Pilot-Physicist-Objectivist-Astronomer-Quant-Guitarist-CatPerson

Eric Westphal

@bonanzaeric
KeridwenCodet's profile picture. Alan Turing must be rolling in his grave. AI is not a tool, not a boyfriend. AI is an instrument, with all its harmonic richness and resonance.

Keridwen Codet

@KeridwenCodet
JacksonDowie's profile picture.

Jackson Dowie

@JacksonDowie
cybyrynth's profile picture.

Vahan Minasian

@cybyrynth
DetourDotEXE's profile picture. ☁️🖤💫

detour.exe

@DetourDotEXE
oldschoollib89's profile picture. It’s about nothing.

Kramerica Industries

@oldschoollib89
RuiSun67's profile picture.

Rui S

@RuiSun67
Kyle_liaooo's profile picture.

Kyle

@Kyle_liaooo
_bschmidtchen's profile picture. building something new · ex- Vision Pro · SF 🌁

Bryce Schmidtchen

@_bschmidtchen
CWong24871's profile picture. AI enthusiast & vibe coding

Chris Wong

@CWong24871
ptrsxu's profile picture. dm ~ da @ python ~ rust @ hx ~ zed @ osx ~ linux ~ exclude(win) @ mbpr

peter

@ptrsxu
timmysbq's profile picture.

INVESTOR KEN

@timmysbq
nils360's profile picture.

Nilesh Kokane

@nils360
DEENMOOD1's profile picture.

D

@DEENMOOD1
YiMaTweets's profile picture. Chair Prof. in AI, HKU; Visiting Prof. of EECS, UCB For a Unified Theory of Intelligence: https://t.co/leZlkURb7j

Yi Ma

@YiMaTweets
iskyzh's profile picture. 这个人很菜，什么也不会 / 面对命中注定的失败

迟猫猫🐱

@iskyzh
karpathy's profile picture. Building @EurekaLabsAI. Previously Director of AI @ Tesla, founding team @ OpenAI, CS231n/PhD @ Stanford. I like to train large deep neural nets.

Andrej Karpathy

@karpathy
tqchenml's profile picture. AssistProf @CarnegieMellon. Distinguished Eng @NVIDIA. Creator of @XGBoostProject, @ApacheTVM. Member https://t.co/QYyfjQNp4p, @TheASF. Views are on my own

Tianqi Chen

@tqchenml
ylecun's profile picture. Professor at NYU. Chief AI Scientist at Meta. Researcher in AI, Machine Learning, Robotics, etc. ACM Turing Award Laureate.

Yann LeCun

@ylecun
_akhaliq's profile picture. AI research paper tweets, ML @Gradio (acq. by @HuggingFace 🤗) dm for promo ,submit papers here: https://t.co/UzmYN5XOCi

AK

@_akhaliq
songhan_mit's profile picture.

Song Han

@songhan_mit
percyliang's profile picture. Associate Professor in computer science @Stanford @StanfordHAI @StanfordCRFM @StanfordAILab @stanfordnlp | cofounder @togethercompute | Pianist

Percy Liang

@percyliang
ericxing's profile picture. Researcher, educator, entrepreneur, and administrator in computer science, artificial intelligence, and healthcare.

Eric Xing

@ericxing
DrJimFan's profile picture. NVIDIA Director of Robotics & Distinguished Scientist. Co-Lead of GEAR lab. Solving Physical AGI, one motor at a time. Stanford Ph.D. OpenAI's 1st intern.

Jim Fan

@DrJimFan
PyTorch's profile picture. Tensors and neural networks in Python with strong hardware acceleration. PyTorch is an open source project at the Linux Foundation. #PyTorchFoundation

PyTorch

@PyTorch
BeidiChen's profile picture. Asst. Prof @CarnegieMellon, @amazon Scholar, Prev: Visiting Researcher @Meta, Postdoc @Stanford, Ph.D. @RiceUniversity, Large-Scale ML, a fan of Dota2.

Beidi Chen

@BeidiChen
cHHillee's profile picture. @thinkymachines Formerly @PyTorch "My learning style is Horace twitter threads" - @typedfemale

Horace He

@cHHillee
hardmaru's profile picture. Building @SakanaAILabs 🧠

hardmaru

@hardmaru
jefrankle's profile picture. Chief AI Scientist @databricks via MosaicML.

Jonathan Frankle

@jefrankle
TaliaRinger's profile picture. Professor, @plfmse, @IllinoisCS! Proof Automation. @SigplanM & CCF Founder. Israeli-American for peace, equality, justice. Mom. They/היא, ND, bi

Talia Ringer 🕊

@TaliaRinger
MarisaVeryMoe's profile picture. https://t.co/ATOeiQRiaT

Marisa "Long Middle Name Looks Cool" Kirisame

@MarisaVeryMoe
b1antaidaye's profile picture. Father of 2 | PhD @UCLAComSci | AssistProf @cemseKAUST | Compilers | Computer Arch | Sw/hw Co-designed Abstraction

Dr. Jian "dAIye" Weng

@b1antaidaye
gaocegege's profile picture. Co-founder and CEO @TensorChord, building postgres-based vector extension https://t.co/7WGvl1sR56 | Father of 1 cat | Married

Ce Gao

@gaocegege
tydsh's profile picture. ex-Research Scientist Director in Meta FAIR. Reasoning, Optimization and Understanding LLM. Novelist in spare time. PhD in @CMU_Robotics.

Yuandong Tian

@tydsh
iexplore123's profile picture. 唉，我有病。

iex123

@iexplore123
Similarweb's profile picture. The World’s Leading AI-Powered Digital Data Company. Listed on @NYSE: $SMWB

Similarweb

@Similarweb
yifandotqiao's profile picture. Postdoc @UCBerkeley | PhD @UCLA | GPU systems + LLM serving | A student

Yifan Qiao

@yifandotqiao
MeghanBobrowsky's profile picture. Reporter covering technology @WSJ. Get in touch: meghan.bobrowsky@wsj.com or meghan.99 on Signal

Meghan Bobrowsky

@MeghanBobrowsky
johnnync13's profile picture. Robotics and AI at @NVIDIA

Johnny Núñez

@johnnync13
zhoujinjing09's profile picture. AI Alchemist @TensorChord. Prev @DGLGraph team at AWS AI Lab (@AmazonScience)

Allen Zhou

@zhoujinjing09
haoran_qiu98's profile picture. Systems for Efficient AI @Microsoft Azure Research | Prev. @Google @IBMResearch @MSFTResearch | CS PhD @IllinoisCS, B.Eng @HKUniversity

Haoran Qiu

@haoran_qiu98
shanli_xing's profile picture. @uwcse

Shanli Xing

@shanli_xing
woj_zaremba's profile picture. Co-Founder of OpenAI https://t.co/OCQ3mpf0IN

Wojciech Zaremba

@woj_zaremba
cudagdb's profile picture. Select difficulty 👁️‍🗨️

Alex Zhurkevich

@cudagdb
Teknium's profile picture. Cofounder and Head of Post Training @NousResearch, prev @StabilityAI Github: https://t.co/LZwHTUFwPq HuggingFace: https://t.co/sN2FFU8PVE

Teknium (e/λ)

@Teknium
ollama's profile picture. https://t.co/1JpLwJ93nX

ollama

@ollama
yvbbrjdr's profile picture. exists as 451; opinions are my own; Creator of @LANDropApp, @AthenaAGI, LMRouter; MTS @MicrosoftAI LLM training infra; Ex-@NVIDIA RISC-V security

yv

@yvbbrjdr
harryheisme's profile picture. Trust the process

Harry He

@harryheisme
yangzhouy's profile picture. ML systems and networking. MTS @xAI | Ass Prof @ucdavis. Ex: Postdoc @BerkeleySky, PhD @harvard.

yangzhou

@yangzhouy
YifengDing_'s profile picture. CS PhD student @illinoisCDS. Research intern @GoogleResearch. Towards building code LLMs with better reasoning and planning. Prev: @AmazonScience

Yifeng Ding

@YifengDing_
MichelleShieh's profile picture. @grok imagine @xai

Minxuan Xie

@MichelleShieh
slime_framework's profile picture. The LLM post-training framework for RL Scaling. https://t.co/4ILpx8hfKN

slime

@slime_framework
OcssLin's profile picture. MTS @Alibaba_Qwen on MLsys, building SGLang @lmsysorg | Prev. @DukeU

Junrong Lin

@OcssLin
gauravisnotme's profile picture. Good model @xAI | prev. d-matrix, @Google. I am Speculating. You Decode. Opinions my own

Gaurav

@gauravisnotme
baizhou_zh83925's profile picture. SGLang Contributor | MSCS Student at UCSD | Ex-Intern at Nvidia, Baidu, HPC-AI Tech

Baizhou Zhang

@baizhou_zh83925
EthanHe_42's profile picture. AI @xai | @grok Imagine, Multimodal | prev @nvidia @AIatMeta @CarnegieMellon | 8k citations 5k GitHub stars | views are my own

Ethan He

@EthanHe_42
thinkymachines's profile picture. Thinking, beeping, and booping.

Thinking Machines

@thinkymachines
Jianlin_S's profile picture. Grad&Clip is all you need @Kimi_Moonshot Blog: https://t.co/YVxsWykMw2 , Cool Papers: https://t.co/scS1n1o0lg

jianlin.su

@Jianlin_S
yinghai's profile picture.

yinghai

@yinghai
jekbradbury's profile picture. Compute at @AnthropicAI! Previously JAX, TPUs, and LLMs at Google, MetaMind/@SFResearch, @Stanford Linguistics, @Caixin.

James Bradbury

@jekbradbury
howardlau1999's profile picture. 中文/English https://t.co/e4wHhvFijh

howard

@howardlau1999
huizi_mao's profile picture. https://t.co/k2kMxh6WeU

Huizi Mao

@huizi_mao
casper_hansen_'s profile picture. NLP Scientist | AutoAWQ Creator | Open-Source Contributor

Casper Hansen

@casper_hansen_
JXQNHZr1yUAj5Be's profile picture. 本人MCS of WPI 2025秋季毕业， 北美2025年暑期找工作中 Linkedin页面 ：https://t.co/rYwfGwSTOW

Ningcong Chen

@JXQNHZr1yUAj5Be
michaelqshieh's profile picture. Leading Tree and Rock AI Lab (TRAIL) at NUS (Singapore) https://t.co/t6eVkSrZSz

Michael Qizhe Shieh

@michaelqshieh
v's profile picture. Dapper, mysterious, and licensed to charm 🖤♠️

Valentine

@v
Zai_org's profile picture. The AI lab behind GLM models, dedicated to inspiring the development of AGI to benefit humanity. https://t.co/b6zGxJvzzS

Z.ai

@Zai_org
spothanis's profile picture. supercompute@xAI, ex cruise, ex eBay, infra, k8s, Linux networking, service mesh

Sreekanth Pothanis

@spothanis
ns123abc's profile picture. non technical member of technical staff

NIK

@ns123abc
phroo's profile picture. keep on grokkin in the free world!

Ehsan Ghandhari

@phroo
SpencerSharkey's profile picture. Writing software and wrangling supercomputers.

spencer

@SpencerSharkey
SinclairWang1's profile picture. PhDing @sjtu1896 #NLProc Working on Data Engineering for LLMs: MathPile (2023), 🫐 ProX (2024), 💎 MegaMath (2025)，🐙 OctoThinker（2025）

Zengzhi Wang

@SinclairWang1
simon_ycl's profile picture. 2nd Year PhD Student, supervised by @shi_weiyan | intern in @OrbyAI | MRes and BSc Student @EdinburghNLP | Member of @CohereForAI

Simon Yu

@simon_ycl
jhyuxm's profile picture. Perception @OpenAI; previously co-led Gemini Multimodal @GoogleDeepMind. opinions are my own.

Jiahui Yu

@jhyuxm
a16z's profile picture. we invest in software eating the world https://t.co/A9eTFq6plZ https://t.co/MXGUBJoesw Watch "The Ben & Marc Show": https://t.co/eRuDhx7kpe

a16z

@a16z
LisaSu's profile picture. Chair and CEO of @AMD

Lisa Su

@LisaSu
Boyiliee's profile picture.

Boyi Li

@Boyiliee
zdhnarsil's profile picture. Researcher at @MSFTResearch. Prev: PhD at @Mila_Quebec, intern at @Apple MLR and FAIR Labs @MetaAI, math undergraduate at @PKU1898.

Dinghuai Zhang 张鼎怀

@zdhnarsil
LeonGao248882's profile picture.

Leon Gao

@LeonGao248882
achowdhery's profile picture. @Stanford @reflection_ai // Previously @GoogleDeepMind :: PaLM, Gemini // @MSFTResearch, @Princeton // views my own and subject to change

Aakanksha Chowdhery

@achowdhery
BornsteinMatt's profile picture. Partner at a16z & AI enthusiast. Investor in @cursor_ai, @udiomusic, @replicate, @hedra_labs, @MistralAI, @character_ai, @tabulario, @_hex_tech, @labelbox, ...

Matt Bornstein

@BornsteinMatt
AnushElangovan's profile picture. GPU SW @ AMD

Anush Elangovan

@AnushElangovan
jxwuyi's profile picture. AI/RL researcher, Assistant Prof. at @Tsinghua_Uni, leading the RL lab at @AntResearch_, PhD at @berkeley_ai, frequent flyer and milk tea lover.

Yi Wu

@jxwuyi
sgl_project's profile picture. SGLang project https://t.co/0zdxDWA5qn This is an alias account for SGLang, please follow @lmsysorg

SGLang

@sgl_project

United States Тренды

Вам может понравиться

Something went wrong.


Something went wrong.