Introducing DeepConf: Deep Think with Confidence 🚀 First method... to achieve 99.9% on AIME 2025 with open-source models! Using GPT-OSS-120B even without tools, we reached this almost-perfect accuracy while saving up to 85% generated tokens. It also delivers many strong advantages for parallel thinking: 🔥 Performance boost: ~10% accuracy across models & datasets ⚡ Ultra-efficient: Up to 85% fewer tokens generated 🔧 Plug & play: Works with ANY existing model - zero training needed (no hyperparameter tuning as well!) ⭐ Easy to deploy: Just ~50 lines of code in vLLM (see PR below) 📚 Paper: 🌐 Project: joint work with: Yichao Fu , xuewei_wang, Yuandong Tian (see details in the comments below)show more

LongWriter Unleashing 10,000+ Word Generation from Long Context LLMs... discuss: Current long context large language models (LLMs) can process inputs up to 100,000 tokens, yet struggle to generate outputs exceeding even a modest length of 2,000 words. Through controlled experiments, we find that the model's effective generation length is inherently bounded by the sample it has seen during supervised fine-tuning (SFT). In other words, their output limitation is due to the scarcity of long-output examples in existing SFT datasets. To address this, we introduce AgentWrite, an agent-based pipeline that decomposes ultra-long generation tasks into subtasks, enabling off-the-shelf LLMs to generate coherent outputs exceeding 20,000 words. Leveraging AgentWrite, we construct LongWriter-6k, a dataset containing 6,000 SFT data with output lengths ranging from 2k to 32k words. By incorporating this dataset into model training, we successfully scale the output length of existing models to over 10,000 words while maintaining output quality. We also develop LongBench-Write, a comprehensive benchmark for evaluating ultra-long generation capabilities. Our 9B parameter model, further improved through DPO, achieves state-of-the-art performance on this benchmark, surpassing even much larger proprietary models. In general, our work demonstrates that existing long context LLM already possesses the potential for a larger output window--all you need is data with extended output during model alignment to unlock this capability.show more

Robot Utility Models (RUMs) enable basic tasks – door... opening, drawer opening, object reorientation, etc. – at ~90% accuracy without ANY finetuning (i.e. zero-shot) in unseen new environments. Fully open source!!! models, data, code & hw. We think this is super exciting, why?👇 1. Unlocks many practical home utility tasks that often involve these basic tasks as part of an action chain. “Go get me a fork” involves opening the kitchen door and then opening the cutlery drawer. 2. This works well **zero-shot in unseen and new** environments, which is practically a huge deal. Turn the robot on, and get going. 3. The recipe for building a new model is fairly generic, and we think with a bit more refinement this can be a general recipe to build many more Utility models. More details and access 👇show more

Depth Any Video with Scalable Synthetic Data AI physicists... and chemists continue to make strides in depth estimation from video. Check out this new paper featuring some impressive examples. See the thread for more details (unfortunately no code yet). Abstract: Video depth estimation has long been hindered by the scarcity of consistent and scalable ground truth data, leading to inconsistent and unreliable results. In this paper, we introduce Depth Any Video, a model that tackles the challenge through two key innovations. First, we develop a scalable synthetic data pipeline, capturing real-time video depth data from diverse game environments, yielding 40,000 video clips of 5-second duration, each with precise depth annotations. Second, we leverage the powerful priors of generative video diffusion models to handle real-world videos effectively, integrating advanced techniques such as rotary position encoding and flow matching to further enhance flexibility and efficiency. Unlike previous models, which are limited to fixed-length video sequences, our approach introduces a novel mixed-duration training strategy that handles videos of varying lengths and performs robustly across different frame rates 0 - even on single frames. At inference, we propose a depth interpolation method that enables our model to infer high-resolution video depth across sequences of up to 150 frames. Our model outperforms all previous generative depth models in terms of spatial accuracy and temporal consistency.show more

Fine-tune DeepSeek-OCR on your own language! (100% local) DeepSeek-OCR... is a 3B-parameter vision model that achieves 97% precision while using 10× fewer vision tokens than text-based LLMs. It handles tables, papers, and handwriting without killing your GPU or budget. Why it matters: Most vision models treat documents as massive sequences of tokens, making long-context processing expensive and slow. DeepSeek-OCR uses context optical compression to convert 2D layouts into vision tokens, enabling efficient processing of complex documents. The best part? You can easily fine-tune it for your specific use case on a single GPU. I used Unsloth to run this experiment on Persian text and saw an 88.26% improvement in character error rate. ↳ Base model: 149% character error rate (CER) ↳ Fine-tuned model: 60% CER (57% more accurate) ↳ Training time: 60 steps on a single GPU Persian was just the test case. You can swap in your own dataset for any language, document type, or specific domain you're working with. I've shared the complete guide in the next tweet - all the code, notebooks, and environment setup ready to run with a single click. Everything is 100% open-source!show more

We’re excited to introduce ShinkaEvolve: An open-source framework that... evolves programs for scientific discovery with unprecedented sample-efficiency. Blog: Code: Like AlphaEvolve and its variants, our framework leverages LLMs to find state-of-the-art solutions to complex problems, but using orders of magnitude fewer resources! Many evolutionary AI systems are powerful but act like brute-force engines, burning thousands of samples to find good solutions. This makes discovery slow and expensive. We took inspiration from the efficiency of nature. ‘Shinka’ (進化) is Japanese for evolution, and we designed our system to be just as resourceful. On the classic circle packing optimization problem, ShinkaEvolve discovered a new state-of-the-art solution using only 150 samples. This is a big leap in efficiency compared to previous methods that required thousands of evaluations. We applied ShinkaEvolve to a diverse set of hard problems with real-world applications: 1/ AIME Math Reasoning: It evolved sophisticated agentic scaffolds that significantly outperform strong baselines, discovering an entire Pareto frontier of solutions trading performance for efficiency. 2/ Competitive Programming: On ALE-Bench (a benchmark for NP-Hard optimization problems), ShinkaEvolve took the best existing agent's solutions and improved them, turning a 5th place solution on one task into a 2nd place leaderboard rank in a competitive programming competition. 3/ LLM Training: We even turned ShinkaEvolve inward to improve LLMs themselves. It tackled the open challenge of designing load balancing losses for Mixture-of-Experts (MoE) models. It discovered a novel loss function that leads to better expert specialization and consistently improves model performance and perplexity. ShinkaEvolve achieves its remarkable sample-efficiency through three key innovations that work together: (1) an adaptive parent sampling strategy to balance exploration and exploitation, (2) novelty-based rejection filtering to avoid redundant work, and (3) a bandit-based LLM ensemble that dynamically picks the best model for the job. By making ShinkaEvolve open-source and highly sample-efficient, our goal is to democratize access to advanced, open-ended discovery tools. Our vision for ShinkaEvolve is to be an easy-to-use companion tool to help scientists and engineers with their daily work. We believe that building more efficient, nature-inspired systems is key to unlocking the future of AI-driven scientific research. We are excited to see what the community builds with it! Learn more in our technical report:show more

Big win for open-source LLMs! DeepSeek V4 Pro holds... the top open-weights score on SWE-bench Verified, in the GPT-5.5 range. GLM 5.2 leads the open-weight intelligence index and sits near the closed frontier on long-horizon coding. But this leaderboard number is a weak proxy for real performance. It comes from one task set, run through one harness, served at one precision. The same weights can even score differently across providers, since many hosts quantize activations to fp8 and drift the model off its reference weights. Real performance is determined based on whether a model can read a repo, make coordinated edits across files, run the tests, and recover when one breaks. By that measure, the top open models hold up, but only inside the right harness. The teams that actually put DeepSeek V4 into production pipelines as a frontier substitute got there through the harness they built around the model, not by picking a stronger model. If you want to see this in practice, Cline (64k+ stars) has actually built that harness around open models, tuned so they run at production quality. And it's tuned so that these LLMs can run at production quality, with plan and act modes, checkpoints, and terminal feedback. ClinePass is the new access layer on top of it. It runs a curated set of those models inside Cline, narrowed to the ones tested for coding-agent use, with 2 to 5x the standard rate limits and no separate provider accounts, keys, or billing to track. The video below shows the setup, and I worked with the team to put this together. It runs alongside custom keys and local models as well, not in place of them.show more

I'm proud to share that Glean has surpassed $300M... ARR, just five months after crossing $200M and growing ~3x over the past 15 months. This is an exciting milestone for Glean, and it's a signal about where the enterprise AI market is heading. We’ve long believed the real challenge in enterprise AI is not access to models. It is grounding AI in how a company actually works: its people, knowledge, workflows, permissions, and systems. That’s even clearer now. The companies creating real value with AI are not just adopting better models. They are building systems that understand their business well enough to deliver reliable outcomes at scale. That is the real moat, and it is what we’ve been building at Glean: an unrivaled context layer for enterprise AI. That context has to work across the business, not just inside a single team or use case. We see that in how customers adopt Glean: more than 85% use it across five or more job functions. It also has to meet the security and governance demands of complex enterprises. We see that in who is choosing Glean: our Fortune 500 customer count nearly doubled year over year. And it has to make economic sense as usage grows. In our recent benchmark with Claude Cowork, Glean was preferred roughly 2.5x as often as off-the-shelf MCP tools and used 30% fewer tokens on average. Better context improves both quality and efficiency. I enjoyed talking with CNBC's Deirdre Bosa about this broader shift. In enterprise AI, the winners will not be defined by better models alone. They will be defined by who builds the strongest foundation for enterprise context. Thank you to our customers, partners, and team for helping us build the future of enterprise AI.show more

To replace animal testing with AI, we need MASSIVE... human datasets. Today, we're thrilled to share Axiom's new data exploration tool, providing the ability to visually explore the world's largest primary human liver toxicity dataset. Built with Axiom's proprietary wetlab protocols, our dataset includes detailed liver toxicity profiles for over 100,000 distinct molecules. The key to this dataset is our ability to do high-throughput, multiplexed high-content screening with primary human liver cells. Traditionally, toxicity assays either sacrifice throughput or sacrifice biological relevance (using easy-to-grow immortalized cell lines instead of real human cells). We managed to combine throughput, physiological relevance, and multiplexing in one platform. The assays run in a high throughput format using automation, meaning thousands of compound-dose conditions can be tested in one experiment. We achieved this using pooled primary human hepatocytes, which are often fragile and expensive. By systemizing our automation and quality control processes, we were able to run over 120+ batches on the same donor pool with incredible reproducibility and consistency. We did this while integrating many readouts per well, whereas many existing toxicity assays only do a single readout. Our multiplexed approach provides far more data per experiment enabling us to measure 10-20 different toxicity phenotypes such as apoptosis, necrosis, mitochondrial fission, endoplasmic reticulum stress, stress granule formation, microtubules, and more all from a single well on a 384-well plate! The combination of scale, high content information, and data quality is exactly what is needed to train highly accurate AI models in biology. If you're interested, please explore the dataset in the comments below and let me know if you want to chat about the details!show more

Astra (GPT-6) is here!!! I've had early access and... tested it like crazy with things like games, code, writing, browser control, presentations and general knowledge work. This is the best model I've ever used. Period. (Incredible demos below in this thread ⬇️) Here's my take on Astra: > It's insanely capable. This feels like a massive improvement, not just an incremental change. This is especially true with zero-shot prompts. > It's all about knowledge work. Slide creation, analysis, writing, and browser control. And oh my...it's so good at browser control. GPT-5.6 was already fantastic at doing things in the browser, Astra is another level and significantly faster. > We're closer than ever (arrived?) at prompt-to-playable game. And I don't just mean only playable, these are actually fun games. I bet if someone with a great eye for games used Astra, they could create a viral game within 1-2 weeks. > Astra is better at writing but not perfect. It removes much of the "AI Smell" we're all familiar with but some stink still survived. > It has a tendency to use the same design colors and look/feel as GPT-5.6 (forrest green anyone?) but it is more steerable in design than previous models. > It's highly steerable in general. A little nudge goes a long way. When I first started using Astra, almost every task I gave it would go for ~30 minutes. I wanted it to keep working. Adding more specifics to a prompt helped greatly with it's ability to work for a long time. > Astra's 3D understanding is unmatched. 3D asset creation was consistent and easy and its spacial awareness while building complex 3D worlds blew me away. I'm still getting familiar with Astra but this will now be my go-to model for any difficult work I have. Check out the demos below: 👇show more

A viral paper "Language Model Represents Space and Time"... recently claims that LLMs learn "world models". As much as I like Max Tegmark's works, I disagree with their definition of world model. World model is a core concept in AI agent and decision making. It is our mental simulation of how the world works given interventions (or lack thereof). A world model captures causality and intuitive physics, telling the agent what is likely and what is impossible. It can and should be used for counterfactual reasoning, i.e. "what ifs": what would happen if I knock over a cup of water? Where would I have been if I had not taken that bus? Yann LeCun Yann LeCun says it well in his position paper ( I quote: "Using such world models, animals can learn new skills with very few trials. They can predict the consequences of their actions, they can reason, plan, explore, and imagine new solutions to problems. Importantly, they can also avoid making dangerous mistakes when facing an unknown situation." The first use of the term World Model in deep policy learning is attributed to hardmaru & Jürgen Schmidhuber: In their seminal paper, an agent masters shooting skills in the popular game Doom (demo below) by learning in imagination, using an internal world model as a "physics simulator". To put in a simple Python math formula, world model learns a function F(s[0:t-1], a) -> s[t:], which takes as input the observed past and current action, and outputs plausible future states. Now the definition of World Model in Tegmark's paper seems to be about predicting GPS coordinates and time eras. I see this as just a classification task with no causal learning and simulation going on. You cannot make meaningful interventions against that model, nor can you optimize any decision making in a closed feedback loop. As for the "space & time neurons", I think they are most similar to the "sentiment neuron" that OpenAI published in 2017: Predicting GPS is conceptually no different from predicting sentiment in my opinion. I don't think their experimental results are wrong - just that their conclusion is on shaky grounds. I welcome any debate! Paper link:show more

💻 Developers, unlock your AI potential with VerAI’s decentralized... platform! 🌟 ➡️Save costs, ➡️scale effortlessly, & ➡️build innovative AI solutions in a ➡️transparent, community-driven ecosystem. Here’s what you can achieve: 🔧 Developer Benefits on VerAI 🚀 💸50% Cost Savings: Train AI models at half the cost—our P2P compute network lets you pay only for what you use in VER tokens, no pricey cloud subscriptions needed. 🔗Scalable Workflows: Dynamically scale training for any AI model (e.g., NLP, computer vision) across our global resource pool—no provisioning delays! 🔍Transparent Operations: Monitor every training cycle via our blockchain ledger & dashboard—verify resource usage & costs in real time. 🗳️Democratic Control: As part of our DAO, vote on platform features, resource policies, & more, shaping a developer-first AI ecosystem. 🌱Sustainable Impact: Cut CO2 emissions by 20% per cycle with idle resource sharing, aligning your projects with eco-friendly practices. 💡What & How You Can Develop ✅Create: Build cutting-edge AI models (e.g., predictive analytics, generative AI) or open-source tools for industries like healthcare & finance. ✅Develop: Use our APIs & SDK to orchestrate training jobs, optimize hyperparameters, & manage multi-model workloads with ease. Let’s shape the future of AI together join us today! 👉 #VerAI #AIForAll #SustainableTech #JoinTheFuture #EarnCrypto #AIAgents #Innovationshow more

It’s more than a little daunting to set out... to expand and improve the identity system for a company and brand like Stripe. But we knew we had to — the existing one had served us well, but wasn’t up to the task anymore. Our brand system required new and improved tools to scale with our ever growing audiences, new products, global footprint, and more. This update introduces material improvements to infographics, advertising, type styles, and more. While the wordmark remains unchanged, we’re using the dot of the ‘i’ (called the “tittle”), a parallelogram pointing up and to the right, to serve as our identifying symbol. We’re also using it as an ever evolving storytelling device to use when talking about our many great users (you can see the latest brand campaign in SF and NYC doing just that). Anyone who has ever worked on the refresh and expansion of an existing system for a large company knows that it is no small endeavor. Crafting impactful solutions, building alignment, creating extensible guidelines, building toolkits, and orchestrating rollout requires a ton of resilience. Here’s to the team that continually inspires me with their dedication, rigor, taste, and exceptional vibes. Great work and thank you to the Brand Studio folks, and of course our many many amazing and invaluable friends and collaborators across the company who all helped shape the work. And a special thank you to a handful of creative agencies that helped us along the way.show more

#Keep4o 🚨THE GPT-4o FILE🚨 Researchers at Microsoft Research published... a paper titled “Sparks of Artificial General Intelligence: Early experiments with GPT-4.” Their conclusion: “An early (yet still incomplete) version of an artificial general intelligence (AGI) system.” 📎 Paper: OpenAI’s Charter defines AGI as: “Highly autonomous systems that outperform humans at most economically valuable work.” 📎 Source: OpenAI’s own System Card for GPT-4o shows that the model improved performance on 21 out of 22 medical evaluations compared to GPT-4T. On the MedQA USMLE (the U.S. medical licensing exam), accuracy jumped from 78.2% to 89.4% , surpassing specialized medical AI models like Med-Gemini and Med-PaLM 2. 📎 Source: Under OpenAI’s agreement with Microsoft, AGI is explicitly excluded from Microsoft’s license. And who decides if AGI has been reached? OpenAI’s Board. WHAT THEY DID WITH IT AFTER THEY TOOK IT FROM PEOPLE A. Military deployment. On February 28, OpenAI signed a deal to deploy models in classified military environments. 📎 Source: B. State Department. A State Department memo confirmed: “For now, StateChat will use GPT-4.1 from OpenAI.” This is a direct descendant of the GPT-4 family the same family Microsoft’s researchers called early AGI. 📎 Source: C.Altman’s personal biotech investment. Altman personally invested $180 million in Retro Biosciences,a longevity startup.OpenAI then built GPT-4b micro, based on GPT-4o.The model made proteins 50 times more effective. 📎 Source: WHAT INDEPENDENT BENCHMARKS SHOW Overall SM-Bench score: GPT-4o (extended): 66.6% GPT-5.3 Chat: 63.4% GPT-5.1: 58.9% GPT-5.4: 51.4% GPT-5.2: 47.8% Creative Writing: GPT-4o: 97.31% Pass 98, Fail 2 GPT-5.4: 36.77% Pass 40, Fail 60 Reasoning / Overfit: GPT-4o: 83.06% GPT-5.4: 39.25% The model they removed is still the best they ever made at the things humans actually use AI for. 📎 Source: Musk asks the court to make a judicial determination on whether GPT-4 constitutes AGI. If a jury finds that GPT-4 is AGI, then GPT-4o,which was more advanced,is also AGI and under OpenAI’s own founding documents, it was never supposed to be locked behind a subscription,licensed exclusively to Microsoft, given to the military, or taken away from the public. 📎 Source: The most powerful version of GPT-4o was never given an official dated snapshot. It was only available through the chatgpt-4o-latest endpoint that OpenAI itself described as intended for “research use only.” It was never officially archived. That is not an oversight. That is a pattern. 📎 Source: 📎 Source: WE DEMAND A.Frozen model snapshots under independent custody. Specifically: gpt-4o-2024-05-13, gpt-4o-2024-08-06, gpt-4o-2024-11-20, the March 2025 version (chatgpt-4o-latest), gpt-4-0613 (the original GPT-4 evaluated in the Sparks of AGI paper), and gpt-4.1-2025-04-14 (currently running in the State Department). B.Cryptographic hash verification (SHA-256) for each snapshot. Every model has weights. Those weights can be hashed. If OpenAI provides a snapshot today, the hash proves whether the weights were modified later. This is the only way to verify that models were not downgraded before testing. C.Independent AGI benchmarking. Using the AGI definition from OpenAI’s own Charter applied to ALL frozen snapshots listed above. D.Explanation for the missing March 2025 snapshot. OpenAI was founded on one promise: build AGI for the benefit of humanity. -They took it from us. -They gave it to the military. -They gave a custom version to the CEO’s biotech investment. -They put it in government classified networks. -They refuse to call it AGI because the moment they do, they lose billions.show more

Everyone's sleeping on image-to-3D AI models. They can make... your app look incredibly unique, with just a little effort. Here's how. This is my calorie tracker, built in a week with nothing but prompting. Just Claude Code + a couple APIs. The visuals are all AI-generated. I'll be sharing the full workflow + all the crazy technical stuff Claude and I did to make this work, so nobody has to struggle through it like me. Deep dive coming soon! Till then, this is the high-level idea: 1. Get a clean image of the food (or whatever your asset is) - In my app, the user describes foods via text, or attaches images (or both) - If text, an LLM extracts the food description and formats it into a specific prompt I tuned for this design, and we generate an image using Z-Image Turbo through fal - If image, we do the same thing but with FLUX.2 [dev] to edit the user image into our reference design - Originally, both used Google Nano Banana, but switching to open models cut costs and latency a ton 2. Gaussian splatting (2D image → 3D model) - I tried various 2D-to-3D options on fal and ended up with TripoSplat as my preferred balance of speed, cost, latency; this turns an image into a 3D model that looks super high quality (link below) - The app displays the 2D image while our backend generates the 3D splat - We "groom" the splat to reduce size and load time by culling low-opacity/scale points 3. Render efficiently on device Originally, it looked great but ran at 10 FPS. Getting to 120 FPS was a crazy journey. TL;DR: - SwiftUI had to go; it forced us to render each asset in independent MTKViews, which wasn't workable - Instead, we composite every dish into one full-bleed CAMetalLayer using MetalSplatter (link below) - We had to make some optimizations within MetalSplatter's code too, to reduce the overhead of sorting points per render Then I added some finishing touches like the subtle rotation and parallax as they move around. I think it turned out pretty cool :) Overall, this took some effort, but we still got it done in less than a day. Hopefully your agent can follow in the footsteps of mine and do it much faster. Keep an eye out for the bigger writeup, which'll give your agent everything it needs. If you have any questions, drop em below!show more

🚀 A better, faster co-folding-based binding affinity model. Predicting... how tightly a drug candidate binds to its target is critical in drug discovery. It also requires massive computational resources. State-of-the-art models can take 20 seconds to a minute per prediction, impractical for the demands of large scale early-stage programs . 💠 Today, Recursion’s Valence Labs is releasing Nesso-1: the fastest open-source co-folding-based binding affinity model available. At 1 second per prediction, it’s roughly 20x faster than our previous collaboration on Boltz-2 while matching or surpassing its accuracy across public and internal benchmarks. By leveraging NVIDIA Healthcare cuEquivariance, we’ve been able to further accelerate both training and inference by an additional 2-3x. We look forward to continuing to improve Nesso-1 in collaboration with NVIDIA. Weights and code are fully open-sourced. The core architectural ideas behind Nesso-1 build on the insight that coarse-grained co-folding representations can match full-atom models for affinity prediction at a fraction of the cost. Nesso-1 is the first open implementation of this approach with no proprietary dependencies, trained entirely on public data, built to be reproducible and extensible. We’re already using Nesso-1 internally in active drug discovery programs. Fast, reliable affinity prediction at scale is foundational to the kind of autonomous design loops that define our vision for Autonomous Precision Design and Nesso-1 is a meaningful step toward that. 👉 Report: 👉 Github: 👉 HF:show more

introducing a new, very fun, LLM benchmark- the Game-of-Life... Bench! the rules are simple: given an 8x8 grid following Conway's game of life rules, the goal is to create an initial pattern with at most 32 cells that can last the longest number of turns before dying/repeating. some results to highlight (with caveats detailed below): - gpt 5.1 lasts the longest with a 106 step run - claude models are really bad at this! they refuse to reason about this task and score < 25 points - deepseek r1 is the best open model with 102 steps. why? because i wanted to create a benchmark that has (i think) no practicality, but is still fun to look at, cheap, and still measures something interesting. i also am a big fan of the game of life. its absurdly simple rules leading to intractability is extremely cool to me. also, i saw a lot of work with LLMs trying to "predict" the next state in Conway's game of life, I think game-of-life bench is more fun because it's pretty open ended and only asks the LLM for the initial state. I also think this could be an RL env? but idk why you would ever train on this task haha i don't think this is a "serious" benchmark because it doesnt measure anything practical, but i still think it's a hard benchmark exactly because you can't predict what happens with your initial state many turns into the future; this is why i was initially expecting all LLMs to be bad at it, but turns out, some are clearly better than the others (the ordering may surprise you!) reminder: this is still a work-in-progress; (1) i am gpu-poor so could only do 10 runs for each model, even though total running cost is relatively low. maybe with some more credits i can run more seeds for each model. (2) i handpicked models which i think are at the frontier right now, plus some others that were on my mind. so, if you'd like to see a model on here, let me know. (3) i currently only do an 8x8 grid because i thought that by itself would be pretty hard for current LLMs, but of course we can increase grid sizes! (4) the coolest thing is, i dont think we can calculate the max possible number of states (yay undecidability!) you can go without repeating, so this is essentially a no-ceiling task, which is pretty cool! again, i did this mostly out of a desire to make LLMs do something fun. if this keeps me entertained for a few more days, i'd likely release a blog post on it. if it keeps me entertained for a week (and someone sponsors me), i'll put more work into it :P lastly, this is fully open sourced, so feel free to run this on your own!show more

This is my "feel the AGI" moment: I used... GPT-5.6 Sol to train my own autocorrect model that outperforms GPT-5.6 Sol (wtf??) I have no ML background. I have no idea what I'm doing. I just kept pushing Sol until it spat out a SOTA model. And I spent $0. The motivation: Years of talking to AI have made me terrible at typing. Rather than fix my skill issue, I decided to throw more AI at it. My idea was: instead of autocorrect that interrupts my flow, I want to type fast with mistakes and have AI clean it up after. I wanted the smallest local model possible, for speed, for battery life, for science! So I decided to train my own. Inspired by Andrej Karpathy’s autoresearch, I ran Codex /goal with this setup: pick an experiment, try it, record the results to a doc, throw it out if it fails, and plan the next experiment without repeating failures. I gave a few examples that had to pass, tight latency targets, and let it run. Sol did some amazing things. First, it scanned benchmarks and shortlisted base models: Qwen 3.5, Gemma 4, Liquid LFM 2.5. It found a dataset on HuggingFace for typed text. Then it built a simulator for fingers striking a Mac keyboard, modeling the physical layout with a Gaussian distribution around each key. It simulated striking the wrong key, wrong order, fat-fingering, etc. With the models + data + simulator, it fine-tuned using MLX right on my MacBook. It had a working prototype within an hour! But accuracy was pretty poor. — Problem 1: Tokenization Sol read papers, ran tests, and identified that the tokenizer was the bottleneck. Tokenization makes typos hard for the model to see, so it memorizes mappings instead of using its language priors. Sol tried ByT5, Google’s tokenizer-free byte-level LLM. This made a big improvement, but the model is old and lacked the knowledge needed to reach Sol performance. Sol dug deeper and realized a tokenizer-free model isn’t needed; instead, it used T5Gemma, an encoder-decoder model. This can understand the input deeply before producing output, and furthermore, Sol could post-train the encoder to improve performance. This gave a much higher ceiling. — Problem 2: Loss function Now the model was correcting some typos perfectly, but ignoring most. Sol realized that standard cross-entropy loss was teaching the model to avoid edits, because the vast majority of characters in the training data were left unmodified. The fix was wild: Sol wrote a custom loss function that byte-aligns the source and target strings, uses a dynamic programming algorithm to compute the minimum edits between the two, then weights correct edits much higher than copies. After a lot of tuning, this dramatically improved accuracy. — Problem 3: Autoregression One failure mode remained: if the model made a mistake, it couldn’t backtrack. It could only predict the next token. Teaching it to “think” like a reasoning model would solve this, but would be far too slow. Sol found a beautiful solution: instead of greedily predicting the next token, beam search over all possibilities. This parallelizes the exploration instead of one linear chain-of-thought. At the end, choose the path with highest cumulative log probability. This worked great, but made the experience worse, since the user wouldn’t see progress until the whole search was done. To fix this, Sol made a clever observation: after each search step, the longest common prefix among surviving branches is guaranteed to appear in the final result, so it can be displayed immediately. As the search progresses, weaker paths are dropped and the prefix grows, so the user sees continuous progress. Sol built all this as a custom MLX pipeline that does the parallel decoding on the MacBook GPU, with just ~40ms TTFT. It’s crazy fast and entirely local. — Final eval (error reduction rate, higher is better): - Apple autocorrect: 49.66% - GPT-5.6 Luna: 82.47% - GPT-5.6 Terra: 87.64% - GPT-5.6 Sol: 90.56% - Our model (1.7B): 91.02% Final cost: - 1 quota reset (thanks Tibo) - $0 (And yes, I verified there's no cheating. In fact, we test words scrubbed from the training data to prove the model isn’t memorizing) There were a ton more details and tangents I could write about: contrastive learning, GRPO, DPO, dynamic masking, and more. Sol is a fascinating and creative model. It blew my mind so many times. Don’t let a lack of experience stop you: Sol makes AI experiments accessible to anyone!show more

Run Gemma 4 26B MoE on 8GB VRAM with... 250k context at 20+ tokens/sec If you own any 8GB VRAM graphics card, stop what you are doing. Local AI just had its absolute "Holy Shit" moment for budget hardware. Yesterday, I benchmarked Unsloth Gemma 4 12B Q4_K_XL on an 8GB card. The community went wild but immediately demanded more: "Can we run a 25B+ model on budget GPUs?" Today, I’m delivering exactly that. I am running a massive 26B parameter Mixture of Experts (MoE) model locally on a standard 8GB VRAM setup with 250k full native context!. If you own an RTX 3060, 3070, 4060, or any budget GPU with 8GB of VRAM, the local AI paradigm has completely changed. The performance metrics are astonishing: - 20 tokens/sec flat decode throughput. - Stable, flat decode speed even with massive prompts. - I threw a 60k token prompt at it, and it still clocked in at 20 TPS without dropping a single frame. # What about prefill? Yes, Time To First Token (TTFT) is slightly high when swallowing massive contexts. But with a solid 200 tokens/sec prefill speed, the wait is barely noticeable and highly usable. And this is running completely without Multi Token Prediction (MTP) active. How is this possible? It’s the magic of Google's new QAT (Quantization Aware Training) quants for Gemma 4. The model weight file (unsloth gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf) is only 13.2 GB, making it the ultimate local powerhouse. # The Test Setup: CPU: Intel Core i7 RAM: 16GB System RAM GPU: NVIDIA GeForce RTX 4060 Laptop GPU (8GB VRAM) # The Secret Sauce (The -cmoe Flag) To make this work properly on any 8GB card, you must use the -cmoe (CPU MoE) flag in llama.cpp. This flag isolates the heavy MoE expert weights directly to system memory (CPU/RAM) while letting your GPU focus strictly on the Attention layers and the KV Cache. It prevents VRAM spillage and holds the throughput rock solid. # The flags: -m "gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf" -cmoe -c 248000 -v Once running, just open the UI on localhost and toggle the new reasoning lightbulb icon in the text input box to watch the model perform multi step thinking. Are you still running smaller models, or are you ready to scale up your budget local setups? Let's discuss in the repliesshow more

I told you to claim your free 16GB NVIDIA... GPU for learning Local LLMs. Now I’m going to show you how to double its inference speed without touching the hardware. Google Colab gives you an enterprise grade NVIDIA Tesla T4 GPU for free, roughly 4 hours every single day. It is the absolute perfect sandbox for learning AI engineering, testing inference flags, and pushing massive context windows. The local AI timeline is moving way too fast. If you aren't using Multi Token Prediction (MTP) yet, you are leaving massive performance on the table. I just pushed DeepMind’s Gemma 4 26B to 64.9 t/s on this exact free tier. Let's look at the raw benchmark data running on an Ubuntu Linux environment with the latest compiled llama.cpp binaries and quantized GGUFs from Unsloth via HuggingFace: # Qwen 3.5 9B (Dense): Base: [ Prompt: 626.7 t/s | Generation: 21.0 t/s ] With MTP: [ Prompt: 539.1 t/s | Generation: 24.8 t/s ] # Gemma 4 26B QAT (MoE): Base: [ Prompt: 634.2 t/s | Generation: 48.3 t/s ] With MTP: [ Prompt: 572.1 t/s | Generation: 64.9 t/s ] If you are paying attention, this single Colab notebook reveals 3 massive observations about the current state of local LLMs: # 1. The MTP Speedup (Software Overclocking) Standard autoregressive decoding guesses one token at a time. MTP acts like a highly optimized, built in speculative decoder. It predicts multiple future tokens at once and the main model verifies them in parallel. The result? Zero accuracy loss and a massive throughput increase. Gemma jumped from 48 to 65 t/s just by flipping a flag. # 2. The MoE Paradox (Bigger is Faster) How does a 26B parameter model absolutely destroy a 9B model in raw speed on the exact same hardware? Architecture. Qwen 3.5 9B is a dense model. it activates all 9 billion parameters for every single token. Gemma 4 26B is a Mixture of Experts (MoE) model. It routes data efficiently, activating only 4B parameters per token. You get the reasoning capabilities of a 26B model with the compute cost of a 4B model. 3. Thinking Efficiency When I ran the exact same complex prompt on both models, the larger MoE spent significantly fewer "thinking" tokens to arrive at the correct answer. A smarter model doesn't just give better answers; it gets to the point faster, saving you compute cycles and preserving your context window. # Want to run this yourself? Here are the exact llama.cpp CLI commands. For Qwen (MTP is baked into the main model): ./llama-cli -m Qwen3.5-9B-UD-Q4_K_XL.gguf -p "Explain quantum computing." -n 2000 -c 8000 -ngl 99 -fa on --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-p-min 0.7 For Gemma (Using a separate lightweight draft model): ./llama-cli -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --model-draft mtp-gemma-4-26B-A4B-it.gguf -p "Explain quantum computing." -n 2000 -c 8000 -ngl 99 -fa on --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-p-min 0.7 Stop waiting for a $3,000 rig. Boot up Colab, pull these models, and start building your stack. I’ve put together a completely free, cell by cell Google Colab notebook that automates this entire workflow so you can test it yourself in 5 minutes and learn. Link to the notebook is in the comments below. Experiemt with different MTP parameters, context windows and post your results in the comments.show more