Image created with gemini-2.5-flash-image with claude-sonnet-4-5. Image prompt: Photorealistic Times Square at golden hour, every billboard and screen displaying different national flags integrated with circuit patterns and AI symbols, multilingual text in Chinese, Arabic, Spanish, and English about artificial intelligence, the Nasdaq tower showing a glowing digital globe, diverse international crowd of tourists and diplomats looking up, vibrant and bustling atmosphere with a United Nations summit energy.
In real-world development scenarios, GLM-4.6 surpasses GLM-4.5 and reaches near-parity with Claude Sonnet 4, while clearly outperforming other open-source baselines. https://x.com/Zai_org/status/1973034644091392002
Very excited to see the Tinker release! @pcmoritz and I had a chance to experiment with the API. It does a nice job of providing flexibility while abstracting away GPU handling. Here’s a simple example showing how to generate synthetic data and fine tune a text to SQL model.”” / X https://x.com/robertnishihara/status/1973455582603649430
Tinker provides an abstraction layer that is the right one for post-training R&D — it’s the infrastructure I’ve always wanted. I’m excited to see what people build with it. “”Civilization advances by extending the number of important operations which we can perform without”” / X https://x.com/johnschulman2/status/1973450054238347314
A flexible API for fine-tuning LMs – Tinker by @thinkymachines Write a simple CPU-only script, and it runs your exact training loop on distributed GPUs. You can fine-tune open models like Llama and Qwen, up to large MoE (Qwen3-235B-A22B), switching them by changing only one https://x.com/TheTuringPost/status/1973827605448306883
Really excited and proud to see Qwen models are in the first batch of supported models for the tinker service! 🤩 we will continue to release great models to grow research in the community 😎 https://x.com/wzhao_nlp/status/1973603599616974970
I’ve been using Tinker at Redwood Research to RL-train long-context models like Qwen3-32B on difficult AI control tasks – specifically teaching models to write unsuspicious backdoors in code similar to the AI control paper. Early stages but seeing some interesting backdoors 👀”” / X https://x.com/ejcgan/status/1973449963259699284
It turns out that the AI jagged frontier worked as a reverse salient, a term from the history of science for a technology or process that holds back the whole system & thus a focus of development. Math & planning were reverse salients, so they have seen the most improvement. https://x.com/emollick/status/1973148208894451908
I had the chance to try @thinkymachines’ Tinker API for the past couple weeks. Some early impressions: Very hackable & lifts a lot of the LLM training burden, a great fit for researchers who want to focus on algs + data, not infra. My research is in RL, and many RL fine-tuning”” / X https://x.com/tyler_griggs_/status/1973450947218252224
Tinker is cool. If you’re a researcher/developer, tinker dramatically simplifies LLM post-training. You retain 90% of algorithmic creative control (usually related to data, loss function, the algorithm) while tinker handles the hard parts that you usually want to touch much less”” / X https://x.com/karpathy/status/1973468610917179630
🚀With early access to Tinker, we matched full-parameter SFT performance as in Goedel-Prover V2 (32B) (on the same 20% data) using LoRA + 20% of the data. 📊MiniF2F Pass@32 ≈ 81 (20% SFT). Next: full-scale training + RL. This is something that previously took a lot more effort”” / X https://x.com/chijinML/status/1973451597393883451
thinking-machines-lab/tinker-cookbook: Post-training with Tinker https://github.com/thinking-machines-lab/tinker-cookbook
[1 Oct 2025] Thinking Machines’ Tinker: LoRA based LLM fine-tuning API https://x.com/Smol_AI/status/1973622595124863044
Announcing Tinker – Thinking Machines Lab https://thinkingmachines.ai/blog/announcing-tinker/
One interesting “”fundamental”” reason for Tinker today is the rise of MoE. Whereas hackers used to deploy llama3-70B efficiently on one node, modern deployments of MoE models require large multinode deployments for efficiency. The underlying reason? Arithmetic intensity. (1/5) https://x.com/cHHillee/status/1973469947889422539
Very excited to see the Tinker release by @thinkymachines! @robertnishihara and I had a chance to experiment with the API, see https://x.com/pcmoritz/status/1973456462346424641
Tinker – Thinking Machines Lab https://thinkingmachines.ai/tinker/
AI Startup Black Forest Labs Shoots for $4 Billion Valuation https://www.pymnts.com/artificial-intelligence-2/2025/ai-startup-black-forest-labs-shoots-for-4-billion-valuation/
Had fun being in Germany to launch a sovereign cloud offering with SAP and Microsoft; important to us to help governments use our frontier models.”” / X https://x.com/sama/status/1971433413086499044
👋 The all new GLM 4.6 from @Zai_org is available on OpenRouter. – GLM-4.6 achieves comprehensive enhancements across multiple domains, including real-world coding, long-context processing, reasoning, searching, writing, and agentic applications. – Context length: increased https://x.com/OpenRouterAI/status/1973037695774384352
Introducing GLM-4.6: Advanced Agentic, Reasoning and Coding Capabilities As our new flagship model, GLM-4.6 brings significant advancements across real-world coding, long-context processing (up to 200K tokens), reasoning, search, writing, and agentic applications. API: https://x.com/Zai_org/status/1973034639708344767
🔎 @Zai_org just released GLM-4.6, and according to Zhihu contributor toyama nao, it’s not chasing GPT-style fireworks — it’s optimizing for real-world agent use. ⚙️ Key Gains (Fig shown below): • Token efficiency = massive win → Reasoning model: 16K → 9K tokens 🔻 https://x.com/ZhihuFrontier/status/1973447038818762841
GLM-4.6 from @Zai_org is available Cline. 200K context (up from 131k), and completes tasks with 15% fewer tokens than GLM-4.5. >48.6% win rate against frontier models, making it one of the most capable open-source models. Available in Cline & the GLM subscription. https://x.com/cline/status/1973099598903386227
The model releases continue… we’ve just released support for GLM-4.6 under the @Zai_org provider! This improves upon the popular GLM-4.5 model, extending the context window from 128k to 200k and achieving higher scores on benchmarks. Try it out today!”” / X https://x.com/roo_code/status/1973022454298837294
GLM-4.6 实测! 最近都很流行更新一大截然后版本就增加0.1嘛? 直接说结论, 前端能力非常好, 效果甚至达到了炫技的程度, 不过随之而来的是复杂场景可能面临性能问题, 所以需要注意防止模型过度发挥或者偶尔留意下代码设计和架构. 不过 python 能力没看到太多变化, 考虑到 python 就一个测试, https://x.com/karminski3/status/1973353334796140716
GLM 4.6 runs quite fast on an M3 Ultra with mlx-lm even at higher precision. Pretty remarkable that it benchmarks competitive to the just-released Sonnet 4.5. Hope those benchmarks hold-up in day-to-day use. Here’s a run using 5.5 bpw quantized model, generating 5.3k tokens at https://x.com/awnihannun/status/1973063906341114327
GLM-4.6 is out on Hugging Face Longer context window: The context window has been expanded from 128K to 200K tokens, enabling the model to handle more complex agentic tasks Superior coding performance: The model achieves higher scores on code benchmarks and demonstrates better https://x.com/_akhaliq/status/1973068098539593932
Wow – Qwen3-Coder-30b AWQ (4bit) on a single 3090, 115 tokens per second. It just zero-shat Pac-Man. It’s no GLM4.5-Air – but, it runs on a single 3090! https://x.com/QuixiAI/status/1971427136977453184
Unitree CEO Wang Xingxing at a Trade Fair in Hangzhou on Saturday: ⦿ Unitree R1 will become the world’s best-selling humanoid robot next year. ⦿ In the first half of this year, the domestic robot industry grew an average rate of 50% to 100% for Chinese intelligent https://x.com/TheHumanoidHub/status/1973158573317501243
Unitree CEO Wang Xingxing expects R1 to be the world’s best-selling humanoid robot next year. Won’t shock anyone if it happens. The company announced the starting price of $5,900 but even at $12k this will sell like hot cakes https://x.com/TheHumanoidHub/status/1973452915366044096
Anyone who sees this video can instantly grasp the (at least) potential for malicious use. And yet nobody with any power (either in the public or at the corporate level) has anything to say (let alone do) to address it, or even acknowledge it.”” / X https://x.com/TheStalwart/status/1973372434133950665
I get data sovereignty in some cases, but there is just no way for new countries to join the frontier model race as long as scaling (in any sense) matters. There is no sovereign model. You will be dependent on the production of Chinese (or US or French) open models as a base.”” / X https://x.com/emollick/status/1972018517919826099
Alibaba Group & partners unveil MMR1: Revolutionizing multimodal reasoning with less data! MMR1 introduces Variance-Aware Sampling (VAS) for stable RL fine-tuning. Tackles unstable optimization & scarce high-quality data. Releasing massive open datasets (~1.6M CoT, 15k RL QA) & https://x.com/HuggingPapers/status/1971487864807469236
Anthropic to triple international workforce in global AI push https://www.cnbc.com/2025/09/26/anthropic-global-ai-hiring-spree.html
Veo3 high quality on Freepik 😀. Again I use Grok4 +Kimi K2 to write my prompts if you are not using them you are missing big time 👀 Veo3 on Flow refuses to generate the prompt for violating the safety, while the API and Gemini App did not have the problem 🤔🤔🤔 { https://x.com/IamEmily2050/status/1951248894035402973
AMD is using Cline as their coding agent for local models. After testing 20+ models, they found what actually works: > 32GB RAM: Qwen3-Coder 30B (4-bit) > 64GB RAM: Qwen3-Coder 30B (8-bit) > 128GB+ RAM: GLM-4.5-Air 10-minute setup with @lmstudio + Cline, linked below”” / X https://x.com/cline/status/1973035211379310708
SPCT/GRM use for RL confirmed. Also: Terminus, as well as 3.2, are trained with specialist model distillation. Target specialities: «mathematics, competitive programming, general logical reasoning, agentic coding, agentic search». https://x.com/teortaxesTex/status/1972611028442861675
Some relevant details about the RL phase for DeepSeek’s latest models: – finetune separate specialist models with RL for mathematics, competitive programming, general logical reasoning, agentic coding, and agentic search, which are used to generate data for final training – https://x.com/iScienceLuvr/status/1972627894636814585
DeepSeek-V3.1-Terminus launches with improved agentic tool use and reduced language mixing errors | VentureBeat https://venturebeat.com/ai/deepseek-v3-1-terminus-launches-with-improved-agentic-tool-use-and-reduced
DeepSeek has launched V3.2 Exp with their new DeepSeek Sparse Attention (DSA) architecture that claims to reduce the impact of the quadratic scaling of compute with context length We’ve independently benchmarked V3.2 Exp as achieving similar intelligence to DeepSeek V3.1 https://x.com/ArtificialAnlys/status/1973230103854456993
Getting ready to try DeepSeek-V3.2-Exp from @deepseek_ai ? vLLM is here to help! We have verified that it works on H200 machines, and many other hardwares thanks to the hardware plugin mechanism. Check out the recipes https://x.com/vllm_project/status/1972664010702221399
DeepSeek-V3.2 shows: – Chinese chips are rising: Day-0 support for Huawei Ascend & Cambricon; – ML compiler: DeepSeek uses TileLang, letting you write Python → compile to optimized kernels on diverse hardware. E.g., 80 lines of Python can reach 95% of FlashMLA’s (CUDA written https://x.com/Yuchenj_UW/status/1972698599554490856
How does @deepseek_ai Sparse Attention (DSA) work? It has 2 components: the Lightning Indexer and Sparse Multi-Latent Attention (MLA). The indexer keeps a small key cache of 128 per token (vs. 512 for MLA). It scores incoming queries. The top-2048 tokens to pass to Sparse MLA. https://x.com/vllm_project/status/1972617272901644345
As @_xjdr says this is not yet profiled fully but we can provisionally accept that DSA had just enabled 1M contexts for ≈everyone. Some “”prototype””, hah. They always underhype their work. Btw, you can drop “”sparse””. Just DeepSeek Attention, it’s cleaner. https://x.com/teortaxesTex/status/1973209017603465324
DeepSeek V3.2 breakdown 1. Sparse attention via lightning indexer + top_k attention 2. Uses V3.1 Terminus + 1T continued pretraining tokens 3. 5 specialized models (coding, math etc) via RL then distillation for final ckpt 4. GRPO. Reward functions for length penalty, language https://x.com/danielhanchen/status/1972613546119991791
Some notes on DeepSeek’s v3.2’s sparse attention mechanism DSA can be thought of as a noncontiguous sliding window where each token only attends to 2048 other tokens. This means that both memory to load and FLOPs is constant at O(2048) (for decode at least). https://x.com/nrehiew_/status/1973193918662713510
💻 API Update 🎉 Lower costs, same access! 💰 DeepSeek API prices drop 50%+, effective immediately. 🔹 For comparison testing, V3.1-Terminus remains available via a temporary API until Oct 15th, 2025, 15:59 (UTC Time). Details: https://x.com/deepseek_ai/status/1972604777474458033
Official release of DeepSeek-V3.2-Exp with DeepSeek Sparse Attention + massive price cuts! DeepSeek Sparse Attention (DSA) makes inference cheaper (especially long-context) by learning which past tokens matter for each new token and running full attention only on those. DSA https://x.com/scaling01/status/1972613330209485282
🚀 Introducing DeepSeek-V3.2-Exp — our latest experimental model! ✨ Built on V3.1-Terminus, it debuts DeepSeek Sparse Attention(DSA) for faster, more efficient training & inference on long context. 👉 Now live on App, Web, and API. 💰 API prices cut by 50%+! 1/n”” / X https://x.com/deepseek_ai/status/1972604768309871061
🛠 Open Source Release 🔗 Model: https://x.com/deepseek_ai/status/1972604780469239843
DeepSeek-V3.2-Exp/DeepSeek_V3_2.pdf at main · deepseek-ai/DeepSeek-V3.2-Exp https://github.com/deepseek-ai/DeepSeek-V3.2-Exp/blob/main/DeepSeek_V3_2.pdf
deepseek-ai/DeepSeek-V3.2-Exp · Hugging Face https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp
Ollama’s cloud models just got an update! ollama run kimi-k2:1t-cloud Welcome Kimi K2 by @Kimi_Moonshot for the 1 trillion parameter model! ollama run deepseek-v3.1:671b-cloud DeepSeek V3.1 Terminus by @deepseek_ai for the updated updated that is better in language”” / X https://x.com/ollama/status/1971750071483167010
Thank you @NVIDIADC for supporting vLLM for @deepseek_ai model launch. Blackwell is now the go to release platform for new MoEs, and we cannot do it without the amazing team from NVIDIA.”” / X https://x.com/vllm_project/status/1973238773090754735
My entry for @AgentSploots contest. Choose @Vader_AI_ and learned so much creating this short video Used a combo of Suno, Veo3, Kling, Grok, and AI in Photoshop. Was very impressed where AI Music is at with Suno and it’s Song editing tools One day a @virtuals_io agent will https://x.com/JDHyper/status/1949962624508084439
Daiwa Securities is hiring startup Sakana AI to build an AI tool analyzing investor profiles, joining other firms adopting the technology (Bloomberg: https://x.com/SakanaAILabs/status/1974109165623853365
We are pleased to announce our partnership with Daiwa Securities, a major financial services firm in Japan. https://x.com/SakanaAILabs/status/1973935631354245286
📢 New Model Drop: HunyuanImage 3 is now on Yupp! It’s @TencentHunyuan’s latest open-source text-to-image model, internally developed from their native, multimodal LLM. We gave it a spin with some prompts: https://x.com/yupp_ai/status/1972449455627727056
Tokyo-based robotics startup Telexistence is partnering with 7-Eleven Japan to develop the humanoid robot “”Astra.”” The initiative will accelerate dataset development for VLA models. Expected in stores by 2029, Astra aims to automate routine tasks. Japan has 20,000 7-Eleven https://x.com/TheHumanoidHub/status/1973523706925424825
The Secret Power of Global Load Balancing in MoE Models New research reveals how a simple change to load balancing in MoE architecture delivers remarkable performance gains 📈 Traditional frameworks calculate Load Balancing Loss (LBL) within each micro-batch, forcing even https://x.com/daddyofadoggy/status/1973759113554174251
A senior government official of the UAE, Abdulla M. Alhamed, met Optimus and Elon at Tesla HQ in California. https://x.com/TheHumanoidHub/status/1972093872177401983
YouWare now supports GLM-4.6 @Zai_org Clear gains vs GLM-4.5: +27% in coding performance Lower credit usage Try it now. https://x.com/YouWareAI/status/1972990785428811923
📢 New Model Drop: GLM 4.6 is now on Yupp! This new model from @Zai_org offers enhanced coding, long-context processing, reasoning, and writing. We tried it out with some prompts: https://x.com/yupp_ai/status/1972994220144427266
We released a patch this week (v3.32.3) which includes GLM-4.6. The model hits 94.9% success rate on diff edits in our testing. That’s within 1.3 points of Sonnet 4.5 at 10% the cost.”” / X https://x.com/cline/status/1973870619013136850
New: see the most popular LLMs for different languages on OpenRouter Here’s @xai’s Grok 4 Fast dominating German prompts and completions: https://x.com/OpenRouterAI/status/1974122770645864767
Alibaba bets big on AI with Nvidia tie-up, new data center plans https://interestingengineering.com/culture/alibaba-nvidia-ai-partnership-expansion
More to come soon https://x.com/AravSrinivas/status/1973255116582166733
People often ask about the relationships between the different Qwen models and the logic behind their naming. Currently, we have several distinct model families: LLM, Coder, VL, Omni, and a newer addition, Image. While our long-term goal is to unify them into a single, truly”” / X https://x.com/JustinLin610/status/1973974975976808808
Qwen-Image-2509 improves consistency ✨ Hope you enjoy it! 🤝 Thanks @drawthingsapp for showcasing it so well!”” / X https://x.com/Alibaba_Qwen/status/1973668568412856595
Prime-rl has now extensive support for MoE both for RL and SFT, we have been training 100B+ model with it We have support for: * Qwen3 a3-30b * GLM series and Moonlight * adding gpt oss series as we speak we end up rewriting most of the modelling code to make it works with”” / X https://x.com/samsja19/status/1973624615768674612
Proud to see Qwen in the first wave of supported models — and thrilled to empower researchers and developers with flexible, accessible fine-tuning tools. We’ll continue releasing powerful open models to support research, innovation, and collaboration across the community. 💡🤝 https://x.com/Alibaba_Qwen/status/1973665010615218421
still experimenting with LoRA based on the @thinkymachines configuration and just implemented it in colab. In this notebook I set up a fine tune of Qwen/Qwen3-0.6B on the OpenR1-Math dataset with lora rank of 1. with this setup you can get the same reward accuracy as full https://x.com/ben_burtenshaw/status/1973776491843297386
21tok/s for qwen 30b/a3b on cpu, 4tok/s on qwen 232b moe too. MoE might be like.. excellent for CPU inference? About to try GLM 4.6″” / X https://x.com/Teknium1/status/1974039942751006816
💫 Introducing RLP: Reinforcement Learning Pretraining—information-driven, verifier-free objective that teaches models to think before they predict 🔥+19% vs BASE on Qwen3-1.7B 🚀+35% vs BASE on Nemotron-Nano-12B 📄Paper: https://x.com/shrimai_/status/1973113867455832096
You can just run qwen3-coder on a macbook w/ @lmstudio”” / X https://x.com/cline/status/1971591597080064121
Qwen3 VL 235B works surprisingly well and is 10x cheaper than Sonnet”” / X https://x.com/scaling01/status/1973777774121984175
🎨Qwen-Image-Edit-2509 just redefined architectural visualization! 🏙️From sketch to stunning reality — seamlessly transfer ANY style onto buildings & cityscapes. → Traditional Chinese patterns? Done. → Photorealistic urban render? Perfect. AI doesn’t just edit images — it https://x.com/Alibaba_Qwen/status/1972558448626987264
Just dropped FP8-quantized version of Qwen/Qwen3-VL-235B-A22B-Instruct. It cuts disk size and GPU memory needs by about 50%, while maintaining >99.6% accuracy. https://x.com/RedHat_AI/status/1973932224400798163
Cool post on integrating Qwen3 Coder + Cline + LM Studio for high-quality local coding: https://x.com/awnihannun/status/1971603427131351218
