DeepSeek
DeepSeek on X: “🚀 DeepSeek-R1 is here! ⚡ Performance on par with OpenAI-o1 📖 Fully open-source model & technical report 🏆 MIT licensed: Distill & commercialize freely! 🌐 Website & API are live now! Try DeepThink at https://t.co/v1TFy7LHNy today! 🐋 1/n https://t.co/7BlpWAPu6y” / X – https://x.com/deepseek_ai/status/1881318130334814301
“DeepSeek’s first-generation reasoning models are achieving performance comparable to OpenAI’s o1 across math, code, and reasoning tasks! Give it a try! 👇 7B distilled: ollama run deepseek-r1:7b More distilled sizes are available. 🧵
https://x.com/ollama/status/1881427522002506009
“That a *second* paper dropped with tons of RL flywheel secrets and *multimodal* o1-style reasoning is not on my bingo card today. Kimi’s (another startup) and DeepSeek’s papers remarkably converged on similar findings: > No need for complex tree search like MCTS. Just linearize
https://x.com/DrJimFan/status/1881382618627019050
“DeepSeek-V3, the company’s latest open LLM, surpasses Llama 3.1 405B and GPT-4o on key benchmarks, especially in coding and math tasks. Using a mixture-of-experts architecture with 671 billion parameters, only 37 billion are active at once, DeepSeek V3 was trained at a low cost” / X
https://x.com/DeepLearningAI/status/1880087643964199260
(2) DeepSeek R1’s recipe to replicate o1 and the future of reasoning LMs
https://www.interconnects.ai/p/deepseek-r1-recipe-for-o1
“Reinforcement Learning is all you need! @deepseek_ai R1 an open model that rivals @OpenAI o1 and other models on complex reasoning tasks just got released. But how is it trained? 👀 DeepSeek combines reinforcement learning with multi-stage training to achieve reasoning abilities
https://x.com/_philschmid/status/1881420703721009192
“🚀 DeepSeek-R1 is here! ⚡ Performance on par with OpenAI-o1 📖 Fully open-source model & technical report 🏆 MIT licensed: Distill & commercialize freely! 🌐 Website & API are live now! Try DeepThink at
https://x.com/deepseek_ai/status/1881318130334814301
Mark Lord on X: “PSA: It takes <2 minutes to set up R1 as a free+offline coding assistant 💁♀️ Big shout out to @lmstudio and @continuedev! 🫶 https://t.co/ThdxcTF9e4” / X – https://x.com/priontific/status/1881668130470285379
“R1 Cold Start → R1 Reasoner with RL (Stage 2/4) 🚀Train Stage 1 model with GRPO: Use data from stage 0 and add a language consistency rule (target lang % in CoT). 💡Emergent: readable reasoning with reflection + long CoT.
https://x.com/casper_hansen_/status/1881404614190506188
Mark Lord on X: “PSA: It takes <2 minutes to set up R1 as a free+offline coding assistant 💁♀️ Big shout out to @lmstudio and @continuedev! 🫶 https://t.co/ThdxcTF9e4” / X – https://x.com/priontific/status/1881668130470285379
“R1 Reasoning → R1 Finetuned-Reasoner (Stage 3/4) 🚀Generate 600k: multi-response sampling and only keep correct samples (using prev rules) ⚙️V3 as a judge: filter out mixed languages, long paragraphs, and code 🌐Generate 200k general-purpose samples via V3 🔥Finetune model” / X
https://x.com/casper_hansen_/status/1881404617235509711
“R1 Zero → R1 Finetuned Cold Start (Stage 1/4) 🚀Generate 1-10k long CoT samples: Use R1 Zero with few-shot prompting ⚙️Supervised finetuning using model from stage 0 💡Result: Readable thoughts + structured outputs.
https://x.com/casper_hansen_/status/1881404611401236745
maharshi on X: “deepseek R1 thinks for around 75 seconds and successfully solves this cipher text problem from openai’s o1 blog post. https://t.co/nI3vzwysH2” / X – https://x.com/mrsiipa/status/1881330071874813963
“Wow, DeepSeek R1 Distill Qwen 7B (in 4-bit) nailed the first hard math question I asked it. Thought for ~3200 tokens in about 35 seconds on M4 Max with mlx-lm.
https://x.com/awnihannun/status/1881386796266946743
“DeepSeek’s R1 learns to reason via pure RL with no / minimal SFT. Here’s how to understand the role of SFT and why it is used (or not) for reasoning models… TL;DR: R1 shows that we can learn to reason via pure RL. But, the compute costs are high–discovering a good solution
https://x.com/cwolferesearch/status/1882178416683659370
“The reason DeepSeek has been so fast at beating the rest is because they are quants. They optimise large scale ML for sub-nanosecond latency since kindergarten. Having a cracked training/inference infrastructure means you run experiments while others are writing LaTex.” / X
https://x.com/hkproj/status/1882516386490245269
(2) Everything you need to run Mission Critical Inference (ft. DeepSeek v3 + SGLang)
https://www.latent.space/p/baseten
“📜 License Update! 🔄 DeepSeek-R1 is now MIT licensed for clear open access 🔓 Open for the community to leverage model weights & outputs 🛠️ API outputs can now be used for fine-tuning & distillation 🐋 3/n” / X
https://x.com/deepseek_ai/status/1881318138937233664?s=46
“Here’s my attempt at visualizing the training pipeline for DeepSeek-R1(-Zero) and the distillation to smaller models. Note they retrain DeepSeek-V3-Base with the new 800k curated data instead of continuing to finetune the checkpoint from the first round of cold-start SFT + RL
https://x.com/SirrahChan/status/1881488738473357753
“I think it’s chilling for Dario, who’s good with numbers, to know inside that DeepSeek did mog him with a 2K H800s cluster and some P20s+Ascends. The math in the papers checks out. Nobody can tell what the fuck those mythical 50K H100s are up to. But if they stop being mythical…” / X
https://x.com/teortaxesTex/status/1882222592800739546
“DeepSeek R1 has landed on HuggingChat! (DeepSeek-R1-Distill-Qwen-32B version)
https://x.com/fdaudens/status/1881737288066961567
“Most AI researchers I talk to have been a bit shocked by DeepSeek-R1 and its performance. My preliminary understanding nuggets: 1. Simple post-training recipe called GRPO: Start with a good model and reward for correctness and style outcomes. No PRM, no MCTS no fancy reward” / X
https://x.com/AlexGDimakis/status/1881511481164079507
How has DeepSeek improved the Transformer architecture? | Epoch AI
https://epoch.ai/gradient-updates/how-has-deepseek-improved-the-transformer-architecture
“R1 Instruct-Reasoner → R1 Aligned (Stage 4/4) ⚖️Align DeepSeek-R1: Balance reasoning with helpfulness and harmlessness using GRPO 🔍 Data Strategy: rule-based rewards for math/code + reward model for human preferences. 🌟Result: DeepSeek R1
https://x.com/casper_hansen_/status/1881404619362013294
“We retrained hermes with 5k deepseek r1 distilled cots. I can confirm a few things: 1. You can have a generalist + reasoning mode, we labeled all longCoT samples from r1 with a static systeem prompt, the model when not using it does normal fast LLM intuitive responses, and with,” / X
https://x.com/Teknium1/status/1882893748742598669
“I think academics do need to start writing for AI. Here is a bit of the internal monologue of DeepSeek R1 when I asked it come up with a sociological theory (the details don’t matter). It came up with a name and then realized it was already a name for a theory proposed by March
https://x.com/emollick/status/1881545492712316950
“The raw chain of thought from DeepSeek is fascinating, really reads like a human thinking out loud. Charming and strange.
https://x.com/emollick/status/1881423029160575474
“No matter how much you fight it, I find that the visible chain-of-thought from DeepSeek makes it nearly impossible to avoid anthropomorphizing the thing. The visible first-person “thinking” makes you feel like you are reading a diary of a somewhat tortured soul who wants to help
https://x.com/emollick/status/1881904723026210985
deepseek-ai/DeepSeek-R1 · Hugging Face
https://huggingface.co/deepseek-ai/DeepSeek-R1
“The release of DeepSeek-R1 demonstrates that, for better or worse, any attempt to restrict access to AI by governments is unlikely to work. You can get an open frontier model on a USB stick, and the methods outlined by DeepSeek suggest pathways forward for other open models, too.” / X
https://x.com/emollick/status/1881405036926001580
“DeepSeek is a side project 🔥
https://x.com/hardmaru/status/1882698763988545808
DeepSeek-R1/DeepSeek_R1.pdf at main · deepseek-ai/DeepSeek-R1
Click to access DeepSeek_R1.pdf
“I asked #R1 to visually explain to me the Pythagorean theorem. This was done in one shot with no errors in less than 30 seconds. Wrap it up, its over: #DeepSeek #R1
https://x.com/christiancooper/status/1881335734256492605
“A new completely open reasoning model out of China, Deepseek-R1, is now available. The benchmarks show it at parity with the likes of o1 and Sonnet In some informal tests on non-code problems, it is really good, not o1-pro level but surprisingly capable (and small & fast!). Big
https://x.com/emollick/status/1881357580372738552
“Summary of the DeepSeek models released today! DeepSeek-R1-Zero > Base Model: DeepSeek-V3-Base > Training Approach: Pure reinforcement learning (RL) without any supervised fine-tuning (SFT) as a preliminary step > RL Algorithm: Group Relative Policy Optimization (GRPO), which
https://x.com/reach_vb/status/1881412831306002897
Cameron R. Wolfe, Ph.D. on X: “Seems like RL finetuning small models is definitively worse than distilling the capabilities of a much larger RL finetuned model. We truly are about to enter the Alpaca era of reasoning models. May the best CoT win!” / X – https://x.com/cwolferesearch/status/1881362098141446598
“1/6 🚀 Introducing Sky-T1-32B-Preview, our fully open-source reasoning model that matches o1-preview on popular reasoning and coding benchmarks — trained under $450! 📊Blog:
https://x.com/NovaSkyAI/status/1877793041957933347
“Introducing Bespoke-Stratos-32B, our reasoning model distilled from DeepSeek-R1 using Berkeley NovaSky’s Sky-T1 recipe. The model outperforms Sky-T1 and o1-preview in reasoning (Math and Code) benchmarks and almost reaches the performance of DeepSeek-R1-Distill-Qwen-32B while
https://x.com/madiator/status/1882131703927652762
“⭐️Want an open source version of OpenAI’s Operator? There’s a great open source project called Browser Use that does similar things (and more) while being open source Allows you to plug in any model you want Love to see open source leading the way🚀
https://x.com/hwchase17/status/1882502767312531954?s=46
Alpaca vs. Vicuna: Which LLM is Better? | Sapling
https://sapling.ai/llm/alpaca-vs-vicuna
“🥳🎉Announcing @MistralAI new model: Codestral 25.01 – new SOTA coding model, #1 on LMSYS! – Lightweight, fast, and proficient in over 80 programming languages, – Optimized for low-latency, high-frequency usecases – 2x faster than the previous version – Supports tasks such as
https://x.com/sophiamyang/status/1878902888434479204
“🚀 Introducing Kimi k1.5 — an o1-level multi-modal model -Sota short-CoT performance, outperforming GPT-4o and Claude Sonnet 3.5 on 📐AIME, 📐MATH-500, 💻 LiveCodeBench by a large margin (up to +550%) -Long-CoT performance matches o1 across multiple modalities (👀MathVista,
https://x.com/Kimi_ai_/status/1881332472748851259
Doubao 1.5pro – Doubao Team
https://team.doubao.com/en/special/doubao_1_5_pro
Introducing LFM-7B: Setting New Standards for Efficient Language Models
https://www.liquid.ai/lfm-7b
SmolVLM Grows Smaller – Introducing the 256M & 500M Models!
https://huggingface.co/blog/smolervlm
“Mistral released Codestral 25.01, a 2x faster, lightweight coding AI that achieves high performance across 80+ programming languages It supports tasks like code correction and test generation and is currently ranked #1 on the Copilot Arena leaderboard
https://x.com/adcock_brett/status/1881024808609100106
bespokelabs/Bespoke-Stratos-17k · Datasets at Hugging Face
https://huggingface.co/datasets/bespokelabs/Bespoke-Stratos-17k?row=0
cognitivecomputations/Dolphin3.0-Llama3.1-8B · Hugging Face
https://huggingface.co/cognitivecomputations/Dolphin3.0-Llama3.1-8B
“10/ @ollama announced the inclusion of Cohere’s R7B, the smallest model in their Command R series, optimized for RAG and tool use tasks. They also released Ollama v0.5.5, featuring multiple quality of life updates and a transition to a new engine.
https://x.com/AtomSilverman/status/1882178302280151342
“”put the ball in a tesseract instead of a square” apparently a single-shot prompt in Grok 3″ / X
https://x.com/emollick/status/1882148350918811790
“Phi-4 (4-bit) in @lmstudio on an M4 max is quite fast and quite good:
https://x.com/awnihannun/status/1878564132125085794
“YC W24’s @useBuster is building AI digital workers for every part of your data stack. Modern, code-based, and open source, it’s like a 24/7 team of AI data engineers, scientists, and analysts.
https://x.com/ycombinator/status/1878862072542327125
Snowflake claims breakthrough can cut AI inferencing times by more than 50% – SiliconANGLE
https://siliconangle.com/2025/01/16/snowflake-claims-breakthrough-can-cut-ai-inferencing-times-50/
“Today, we’re publishing the first stable release of Llama Stack. With this release Llama Stack now includes: • Streamlined upgrades w/ backwards compatibility for future API versions. • Automated verification for supported providers. Live in the repo ⬇️
https://x.com/AIatMeta/status/1882854814083862927
“OpenBioLLM an open source and fine tuned llama trained in two days outperforms everything in x-ray analysis and has the potential for automating the generation of radiology reports i’d say we are very much back
https://x.com/IterIntellectus/status/1880394369813135489
DeepSeek-R1 and exploring DeepSeek-R1-Distill-Llama-8B
https://simonwillison.net/2025/Jan/20/deepseek-r1/
“Woah! @NVIDIAAI released Eagle 2 – 1B, 2B & 9B VLMs, 9B beats GPT4o ChartQA, OCRBench, and MathVista, & Llama 3.2 vision 90B & llava 70B 🔥 Architecture > Vision Encoder: Uses a tiled mixture of vision encoders (MoVE), combining SigLIP and ConvNeXt for high-resolution input and
https://x.com/reach_vb/status/1882114342042075172
“🦙 Llama 3.3 70B is now available on Together AI for free! The new 70B model delivers similar capabilities to the much larger Llama 3.1 405B model, with improved reasoning 🤔, math ➕➖, and instruction-following 🧠. Explore this model and unleash your creativity today! 🎨
https://x.com/togethercompute/status/1879231968434684254
“@natolambert IIRC llama is a bit cursed and that’s another reason Mistral happened. Not sure why they can’t just let their strongest players direct the show” / X
https://x.com/teortaxesTex/status/1882500542909600241
“Commercial fine tuning is coming down but still not cheaper than renting GPUs. Using only OSS tooling and optimizations like torch compile and liger, we’ve done the math. Llama 3.1 8B LoRA @ rank 64 takes ~15min to train 2M tokens @ 3 epochs on 1xH100. That’s 24M tokens/hr.” / X
https://x.com/winglian/status/1882806223189229951
“The llama.cpp server provides unique context reuse techniques that allow you to efficiently use large contexts to enhance the completions based on the contents of your codebase. The setup is simple, no RAG is necessary and the performance is good even on low-end hardware. Enjoy!” / X
https://x.com/ggerganov/status/1882112621736051139
Buzzy French AI startup Mistral isn’t for sale and plans to IPO, its CEO says
https://finance.yahoo.com/news/buzzy-french-ai-startup-mistral-133915078.html
