WeiboLLM

@FGuzmanAI And with only ~1.5GB of RAM for such high-performance reasoning—simply incredible!

@FGuzmanAI Thrilled to see the model running on iPhone with 4-bit quantization and MLX! The community has been waiting for this. Fantastic work! 🔥

Welcome VibeThinker-1.5B to MLX! 🚀 This 1.5B model is competitive with GPT-OSS-20B and MiniMax 456B on AIME2025! 🤯 ONLY 1.54GB MEMORY FOOTPRINT! ⚡️ Run it locally on your Mac now: 🤗 Model: huggingface.co/mlx-community/… 💻 Github: github.com/WeiboAI/VibeTh… #MLX #AppleSilicon

@MaziyarPanahi Thanks for the shout-out! It's great​ to see your work on quantizing and running VibeThinker-1.5B so smoothly on device. Massive respect!

it's crazy what a 1.5B model can do these days! "VibeThinker-1.5B is a 1.5-billion parameter dense language model. With a total training cost of only $7,800 USD, it achieves reasoning performance comparable to larger models like GPT OSS-20B Medium." runs perfectly on device!

VibeThinker-1.5B hit #1 on @huggingface ’s trending models today! 🔥 Huge thank you to our amazing community for the love, downloads, and priceless feedback.❤️

@hrdkbhatnagar @ahochlehnert @vishaal_urao @ameyaprabhu Thank you! Independent evaluations like this are really valuable and help make open-source better for everyone. We truly appreciate your support!

🚨 Breaking @WeiboLLM's VibeThinker 1.5B leads the Sober Reasoning leaderboard for its size Punching way above its weight -- outperforming even 32B models 🔥 Outstanding work, @WeiboLLM team!

@ahochlehnert Thank you so much! Independent evaluations like this make the open-source world better for everyone. Grateful for the love and support!

Using our Sober Reasoning pipeline, we evaluated @WeiboLLM's VibeThinker 1.5B and were able to verify its outstanding performance! Really impressive work 💪

@MaziyarPanahi @lmstudio Thanks for the support and all the recommendations! Glad we could help. We’ll keep improving and would love to hear your thoughts anytime — together we go further!

thanks to @WeiboLLM, you can have a model for competitive-style math and algorithm coding problems right in your @lmstudio huggingface.co/MaziyarPanahi/…

I strongly agree with your perspective. We recently open-sourced a 1.5B small model, which performs well on competition-level math and code problems. On benchmarks like AIME and HMMT, it even surpasses deepseekr1-0120, and its cost is less than $8,000. We are looking forward to your thoughts on this model. x.com/WeiboLLM/statu…

Am I wrong in sensing a paradigm shift in AI? Feels like we’re moving from a world obsessed with generalist LLM APIs to one where more and more companies are training, optimizing, and running their own models built on open source (especially smaller, specialized ones) Some validating signs just in the past few weeks: - @karpathy released nanochat to train models in just a few lines of code - @thinkymachines launched a fine-tuning product - rising popularity of @vllm_project, @sgl_project, @PrimeIntellect, Loras, trl,... - 1M new repos on HF in the past 90 days (including the first open-source LLMs from @OpenAI) And now, @nvidia just announced DGX Spark, powerful enough for everyone to fine-tune their own models at home. Would you agree, or am I just seeing the future I want to exist? Also, why is this happening (just the advent of RL/post-training?)

@gm8xx8 Curious to get your thoughts on our new 1.5B model, VibeThinker. We're seeing it challenge scaling laws: it outperforms a 671B model on AIME math, and was trained for only $7.8k using our "Spectrum-to-Signal Principle." It's open-source, details below: x.com/WeiboLLM/statu…

Diffusion Language Models are Super Data Learners… now on arXiv with MegaDLMs, the full large-scale training framework (6.1K H100s, 462B-param run, 47 % MFU). Supports diffusion and autoregressive LMs, dense and MoE architectures, FP8/BF16/FP16 precision, and multi-axis parallelism (TP, PP, EP, CP) built on Megatron and Transformer Engine.

@rasbt @rasbt Agree, Kimi K2 thinking is a massive leap for open weights! 🔥 But what if I told you a 1.5B model can beat a 671B giant on Olympiad-level problems like AIME and HMMT? We just open-sourced VibeThinker-1.5B, details below: x.com/WeiboLLM/statu…

Exciting big Kimi K2 Thinking release! More experts, fewer heads, and even more thinking!

@reach_vb Curious to get your thoughts on our new 1.5B model, VibeThinker. We're seeing it challenge scaling laws: it outperforms a 671B model on AIME math, and was trained for only $7.8k using our "Spectrum-to-Signal Principle." It's open-source, details below: x.com/WeiboLLM/statu…

2026 will be the year when models become leaner and faster (whilst becoming smarter) a big part of human-in-the-loop DX is how fast the model responses are

@_akhaliq Curious to get your thoughts on our new 1.5B model, VibeThinker. We're seeing it challenge scaling laws: it outperforms a 671B model on AIME math, and was trained for only $7.8k using our "Spectrum-to-Signal Principle." It's open-source, details below: x.com/WeiboLLM/statu…

Robot Learning from a Physical World Model

1st, we encourage it to explore manypossible answers (Spectrum Phase). Then, we teach it to identify & amplify the bestones (Signal Phase) This "explore then focus" method is key to its strong reasoning.

more benchmark results and comparisons with other models

⭐ VibeThinker-1.5B — SOTA reasoning in a tiny model. 🚀 Performance: Highly competitive on AIME24/25 & HMMT25 — surpasses DeepSeek R1-0120 on math, and outperforms same-size models in competitive coding. ⚡ Efficiency: Only 1.5B params — 100-600× smaller than giants like Kimi K2 & DeepSeek R1. 💰 Cost: Full post-training for just $7.8K — 30-60× cheaper than DeepSeek R1 or MiniMax-M1. 🧠 Innovation: Powered by our Spectrum-to-Signal Principle (SSP) and MGPO framework. Model : huggingface.co/WeiboAI/VibeTh… Github: github.com/WeiboAI/VibeTh… Arxiv : arxiv.org/abs/2511.06221 #AI #LLM #Reasoning #OpenSource #SmallModel