Stories & updates
Browse the latest LeemerChat write-ups, search for specific topics, and jump into the full articles with one click.
February 22, 2026
Tinker is now generally available. Vision input, Kimi K2 Thinking, and LoRA Without Regret are reshaping what custom model training looks like in 2026. Here's why fine-tuning is more strategically important than ever — and how LeemerLabs Model Foundry is building the infrastructure to prove it.
Tinker is now generally available. Vision input, Kimi K2 Thinking, and LoRA Without Regret are reshaping what custom model training looks like in 2026. Here's why fine-tuning is more strategically important than ever — and how LeemerLabs Model Foundry is building the infrastructure to prove it.
February 12, 2026MiniMax M2.5 launches on LeemerChat with breakthrough performance in Word, Excel, and PowerPoint generation. Scoring 80.2% on SWE-Bench Verified and 76.3% on BrowseComp, M2.5 extends M2.1's coding expertise into general office productivity.
February 11, 2026GLM-5 launches on LeemerChat with major upgrades in scale, training data, and RL infrastructure. Built for long-horizon agentic systems, coding reliability, and complex reasoning under production constraints.
Search and explore everything we have published.
Tinker is now generally available. Vision input, Kimi K2 Thinking, and LoRA Without Regret are reshaping what custom model training looks like in 2026. Here's why fine-tuning is more strategically important than ever — and how LeemerLabs Model Foundry is building the infrastructure to prove it.
MiniMax M2.5 launches on LeemerChat with breakthrough performance in Word, Excel, and PowerPoint generation. Scoring 80.2% on SWE-Bench Verified and 76.3% on BrowseComp, M2.5 extends M2.1's coding expertise into general office productivity.
GLM-5 launches on LeemerChat with major upgrades in scale, training data, and RL infrastructure. Built for long-horizon agentic systems, coding reliability, and complex reasoning under production constraints.
We've integrated Cursor's Cloud Agents API into LeemerChat so you can launch, monitor, stop, and follow-up with autonomous coding agents on your GitHub repos — all from the chat. Just enter your API key in settings and start dispatching agents with natural language.
We've integrated Blackbox Cloud into LeemerChat so you can dispatch autonomous coding agents to your GitHub repos — single-agent or multi-launch — without leaving the conversation. Create, monitor, and cancel tasks with natural language.
What if AI answers came from a council of experts instead of a single voice? KingLeemer orchestrates multiple frontier models to think together, disagree, debate, and converge on answers more reliable than any single model could produce alone.
Kimi K2.5 brings state-of-the-art visual coding, 262K context, and self-directed agent swarms. We’re Ireland’s first AI platform to launch it — and it’s live free on LeemerChat.
Upload a PDF, get it back solved. No prompts required. Introducing our standalone PDF processing workflow with Mistral OCR, Multi-Agent Consensus, and Visual Overlay.
Introducing Codebase Chat with GitHub integration, natural AI Memory for preferences, Second Thought expert consultations, and concurrent generation. The biggest evolution of LeemerChat yet.
Google's Gemini 3 Flash represents a clear shift in how frontier-level AI is delivered in production. Near-Pro-level reasoning and multimodal understanding while remaining fast, responsive, and economical enough for large-scale deployment.
Meet RIN (凛) — a 26B-A3B MoE model running at 450 tokens/second, completely free and unlimited. The precision instrument for builders who value speed over hand-holding. Semi-successor to LeemerGLM.
This holiday season, we're reflecting on an incredible year together. From V3 to V4.9, over 1.5B+ tokens processed, LeemerLite at 1,750 T/s, PowerCode, and welcoming GPT-5.1, Claude 4.5 Sonnet, Gemini 3, and Qwen — here's to building the future together.
Discover how IKEA-inspired design, frosted glass interfaces, and revolutionary durable generation create an AI workspace that feels effortless yet powerful. This is what happens when design meets reliability.
We just dropped LeemerLite: a super-simplified, no-signup chat running gpt-oss-safeguard-20b at world-class speeds. See how it stacks up against GPT-5 Nano, Llama 4 Scout, and Mistral.
Detailed drop of research, email, and automation updates with new model lineup.
A behind-the-scenes look at how we built LeemerGLM on top of Gemma 3 4B, why we paired it with a multimodal specialist, and how it slots into our expert panel.
Vibe coding feels fast, but it hides a $50B cleanup bill. This editorial exposes why 80% of startups crash because of sloppy code and how frontier AI turns vibes into infrastructure.
The AI boom is the biggest technological surge since the internet — and it is dangerously overheated. Based on historical venture cycles, market concentration, and structural economics, 90% of today's AI companies will likely be worth close to zero by 2027. This is not pessimism. It is pattern recognition.
In a world where AI is becoming critical infrastructure, relying on third-party APIs is like renting your brain. We explore why self-hosted, sovereign AI models are the future—and why the smartest companies are already making the switch.
We're launching Ireland's first custom LLM creation studio. Fine-tune frontier models up to 235B parameters using Tinker distributed training, powered by Thinking Machines Lab. Build domain-specific intelligence layers that you own and deploy anywhere.
We tested GPT-5.1, Claude Sonnet 4.5, Grok-4.1-Fast, and Gemini 2.5 Pro across coding, reasoning, writing, vision, research, and speed. The results reveal why using multiple models in one chat is the future of AI.
How LeemerChat used BotID Deep Analysis to shut down coordinated synthetic agents without slowing down real users.
A deep dive into the union model architecture powering Leemer Heavy's iterative research orchestration and Heavy (Fast)'s rapid debate synthesis system.
How we built Deep Research by orchestrating three world-class AI models together, powered by K2-Thinking—the world's strongest reasoning open-source model. Plus a preview of Ultra version running autonomously for 3-4 hours.