Philipp Schmid

Llama 3 released! 🚨🔔 Meta just released their best open LLM! 👑🚀 Llama 3 is the next iteration of Llama with a ~10% relative improvement to its predecessor! 🤯 Llama 3 comes in 2 different sizes 8B and 70B with a new extended tokenizer and commercially permissive license! ✅

𝗡𝗲𝘄 𝗮𝗻𝗱 𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁𝘀 𝘁𝗼 𝘃𝟮✨:
🔠 Trained on 15T Tokens & fine-tuned on 10M human annotated samples
🧮 8B & 70B versions as Instruct and Base
🚀 Llama 3 70B best open LLM on MMLU (> 80 🤯)
🧑🏻‍💻 Instruct good at coding 8B with 62.2 and 70B 81.7 on Human Eval
✍🏻 Tiktoken-based tokenizer with a 128k vocabulary
🪟 8192 default context window (can be increased)
🧠 Used SFT, PPO & DPO for alignment.
💰Commercial use allowed ✅
🤗 Available on Hugging Face
🤝 1-click deployments on Hugging Face, Amazon SageMaker, Google Cloud
🔜 more model sizes & enhanced performance

Blog: https://lnkd.in/ehXXavJ8
Models: https://lnkd.in/ek2pJviv
Chat-Demo: https://lnkd.in/eyRHH2X4

Massive kudos to Meta for continuing its commitment to open AI. Honored to partner with Joseph Spisak and team! 🤗 The gap is melting. 🧊

RAG Developer Attention! 🔔 Docling is a new library from IBM that efficiently parses PDF, DOCX, and PPTX and exports them to Markdown and JSON. It supports advanced PDF understanding and seamless integration with LlamaIndex and LangChain.

TL;DR:
🗂️ Parses numerous document formats (PDF, DOCX, PPTX, and more) into Markdown & JSON.
📑 Advanced PDF processing: handles layout, reading order, and tables.
🧩 Unified document representation for easier processing.
🤖 integration with LlamaIndex and LangChain for RAG applications.
🔍 Includes OCR for scanned PDFs.
💻 User-friendly CLI and Python API.

Docs: https://lnkd.in/ePnM84Fy
Github: https://lnkd.in/eJT46F4j

Google just released Gemini. 🧠 Their most capable and general model. Gemini is a multimodal model trained across image, audio, video, and text data. Based on the Technical Report, the performance of the biggest model (ultra) seems to be on par or slightly better than GPT-4. 💪🏻

𝗧𝗟;𝗗𝗥:
✏️ Support Text, Vision, and audio inputs or outputs, e.g., transcription, image generation
📏 Decoder architecture with 32k context length and Multi Query Attention (MQA)
👀 Visual Encoder inspired by Flamingo
📚 Trained on web documents, books, and code including image, audio, and video data. No details on the number of tokens.
3️⃣ Comes in 3 sizes: Ultra, Pro, and Nano for different use cases
⚡️ Trained using TPUv5e and TPUv4
📱 Pixel 8 Pro is the first smartphone engineered to run Gemini Nano
⬆️ Performance of Gemini Ultra on par or slightly better than GPT-4
💪 Strong capabilities across reasoning, coding, language understanding
🔁 Used RLHF to fine-tune the model
❌ No information about the Size of Ultra and Pro models
❌ No detailed training data

Blog post: https://lnkd.in/eud5Wd74
Technical Report: https://lnkd.in/e5NvbhRv

Are Vector Databases here to stay? 🔍 Yes, it seems LLMs are lost in the Middle and lose focus on long inputs.🗺👁‍🗨
In the “Lost in the Middle: How Language Models Use Long Contexts,” a group of researchers from Stanford tried to understand better how LLMs make use of the context📚✨ Below are some of my key takeaways: 📝

🔍 𝗢𝗯𝗷𝗲𝗰𝘁𝗶𝘃𝗲:
- Analyze and evaluate how LLMs use the context by identifying relevant information within it.

💻 𝗜𝗺𝗽𝗹𝗲𝗺𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻:
- Tested open-source (MPT-30B-Instruct, LongChat-13B(16K)) and closed-source (OpenAI’s GPT-3.5-Turbo and Anthropic’s Claude 1.3) models
- Multi-document question-answering where the context included multiple retrieved documents and one correct answer, which position was shuffled around
- Key-value pair retrieval to analyze if longer contexts impact performance

💡 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴𝘀:
- Best Performance when relevant information is at the beginning
- Performance decreases with an increase in context length
- Too many retrieved documents will harm performance
- Improving the retrieval and prompt creation step with Cross-Encoders (ranking) could potentially boost performance by up to 20%
- Extended-context models (GPT-3.5-Turbo vs. GPT-3.5-Turbo (16K)) are not better if the prompt fits the original context.

Check out the full paper here: https://lnkd.in/etxXnVyp

Combining Retrieval with Ranking should yield the best performance in RAG for Question Answering. 👨‍🔬

Remember that these are just my personal findings. Make sure always to conduct your own research and analysis. 🤗

New Open-source LLMs! 🤯 The Falcon has landed! 🦅 TII just released two new open-source LLMs called Falcon, which comes into size 7B trained on 1.5T tokens and 40B trained on 1T Tokens. 🚀🔥

Falcon:
💥 outperforms comparable open-source models (e.g., MPT-7B, StableLM, RedPajama etc.), thanks to being trained on 1,500B tokens of RefinedWeb enhanced with curated corpora.
🏎 uses FlashAttention and multi-query Attention
🔠 has a 2048 context window
💰 comes with a license allowing commercial use, but with limitations. Make sure to check the license‼️ -> now Apache 2.0
🧠 was trained on Amazon SageMaker, on 384 A100 40GB GPUs in P4d instances.
🌍 40B was trained on a multi-lingual dataset, including German, Spanish, French

Models are available on Hugging Face 🤗
7B: https://lnkd.in/ejpGndA2
40B: https://lnkd.in/e6ESxVTK

Check out the official announcement:
👉 https://falconllm.tii.ae/

Llama can now code!🤯 🔔 @Meta just released CodeLlama, a foundation model for Code Generation. 🧑🏻‍💻 CodeLlama is the next iteration of Metas Llama family and three different sizes, 7B, 13B, and 34B, and has the same commercial-friendly license as Llama 2! 😍

CodeLlama key facts✨:
🧮 7B, 13B & 34B parameter version
🛫 initialized from Llama 2
🔠 Trained on 500B Tokens
🐍  Python version & Instruct version
✅ Commercial use allowed
😕 Training data unknown
🪟 16384 context window
🤗 Available on Hugging Face

Models: https://lnkd.in/e6VtKqGU
Announcement: https://lnkd.in/eaSan-Tt
Paper: https://lnkd.in/e-QTjQym

Big news for any developer or AI enthusiast! 🤗 Transformers just released Agents 🤖 to easily build GenerativeAI applications and autonomous agents using LLMs like OpenAssistant, StarCoder, OpenAI, and more.

📈 With Transformers Agents, you can remove the barrier of entry to machine learning and start building powerful agents that respond to complex queries and offer a chat mode. Plus, it's full-multimodality, allowing you to work with text, images, video, audio, and documents. 🤯

👉 https://lnkd.in/eN7HGFe5

But how does it work in practice? It's as simple as building a prompt:
1️⃣ Tell the agent what it aims to do.
2️⃣ Give it tools.
3️⃣ Show examples.
4️⃣ Give it a task.

Transformers Agents comes with built-in tools like Doc QA, Text QA, Image QA, Speech-to-text and Text-to-speech, Text classification, summarization, translation, download, Image generation, captioning, segmentation, upscaling, and Text to video. And it's designed to be EXTENSIBLE, meaning you can add your own tools or use community-contributed tools. 🤖 📝🎤 🎨🌅 🏷️

Ready to get started now or check out the example notebook: https://lnkd.in/e3XMBBCq

Llama 3 extended to almost 100,000-token context! ✅ By Combining PoSE and continuing pre-training on Llama 3 8B base for 300M tokens, the community (Wing Lian) managed to extend the context from 8k to 64k. 🚀 Applying rope scaling afterward led to a supported context window of close to 100,000 with perfect recall. 🤯🚀

PoSE can extend the context window of LLMs by simulating long inputs using a fixed context window during training. It chunks the document into smaller pieces and simulates them as “long” versions, which significantly reduces memory and time overhead while maintaining performance.

𝗜𝗻𝘀𝗶𝗴𝗵𝘁𝘀:
🚫 Don't increase rope_theta during pertaining
🚀 Rank-stabilized LoRA converged much quicker than regular LoRA
⬆️ Increased the RoPE theta to extend the context to ~90k
➕ Adapters can be merged with any Llama 3 model to extend the context

Llama 3 8B 64k: https://lnkd.in/d6cprxvT
Original Thread: https://lnkd.in/dnVn8vKu
PoSE Paper: https://lnkd.in/dmtDNwwe

Excited to announce “llm-sagemaker” a new Terraform module to easily deploy open LLMs from Hugging Face to Amazon Web Services (AWS) SageMaker real-time endpoints! 👀 Infrastructure as Code (IaC) tools are crucial for moving your AI Applications from Notebooks into production! 🚀

TL;DR:
🚀 New HashiCorp Terraform module simplifies LLM deployment to Amazon SageMaker
🤖 Support for popular models like Meta Llama 3, Mistral AI, Mixtral, and Cohere Command
🛠️ Handles IAM roles, SageMaker Model, Endpoint Configuration, and Autoscaling
⚡ Example deploy Llama 3.1 8B Instruct in just minutes
🔧 Customizable configurations for TGI
💻 Easy integration with AWS SDK for inference
✅ Includes integration tests using Gruntwork terratest

Blog: https://lnkd.in/eQM5KtSD
Module: https://lnkd.in/e8THb3Rd

If you have feature requests, please open an issue. 🤗

Coming in hot!🔥 Another open-source LLM just got released by Stability AI. 🖼 The first release of StableLM includes a 3B and 7B parameter model, with a 15-65B model to follow. 🤯 
Models are released under CC BY-SA license. 📄

Models are available on Hugging Face! 🤗


Model: https://lnkd.in/ehDJ-H5T
RHLF Demo: https://lnkd.in/eYCAzyDY

DoRA a new, better, and faster LoRA? 🤔 DoRA, or Weight-Decomposed Low-Rank Adaptation, is a new, Parameter Efficient fine-tuning technique that claims to enhance the learning capacity and training stability of LoRA, while avoiding any additional overhead. 🤯

DoRA decomposes the weights into two components, magnitude, and direction. DoRA adjusts how important each part of the data is (magnitude) and how the model should focus on learning (direction). It's like fine-tuning the details of a picture without needing to redraw the whole thing. 🖼️

𝗜𝗻𝘀𝗶𝗴𝗵𝘁𝘀:
🏅 DoRA consistently outperforms LoRA
🤗 Supported in Hugging Face PEFT
♻️ Trained Adapters can be merged back into the model
📈 +3.4% on Llama 7B and +1.0% on Llama 13B compared to LoRA on common reasoning
🔐 Improved training stability compared to LoRA
❌ In PEFT, DoRA only supports linear layers at the moment.

Paper: https://lnkd.in/eiQ-ZVTW
Peft documentation: https://lnkd.in/eMYWeGip

Gemma 4 is here! 4️⃣ Our most capable, agentic open model, built on the same research as Gemini 3. Reasoning. Multimodal. Four sizes (2B to 31B). Base + Instruct. ✨

Released under Apache 2.0. Runs on your phone, laptop, or servers. All you need to know about Gemma 4:

4️⃣ 4 sizes (E2B, E4B, 26B4A, 31B)
🪟 Up to 256K context window
🛠️ Native function-calling, structured JSON output
👁️ + audio on edge models (E2B/E4B)
🌍 Trained on 140+ languages
🏆 31B ranks #3 open model on Arena AI
🪪 Apache 2.0 license
1️⃣ Fits on a single GPU
🚀 Gemma E4B == Gemma 3 27B

All versions support native function-calling and structured JSON output to build agents that can run locally. The small models (E2B, E4B) can run entirely offline on mobile supporting vision, audio, everything on-device.

Start building with Gemma 4 now.

Try in Google AI Studio → https://lnkd.in/ddQhAsCs
Hugging Face → https://lnkd.in/dhNBspd5
Kaggle → https://lnkd.in/dffxshtG
or in your favorite ecosystem tool!

Blog → https://lnkd.in/d_EXTGCn

What if one embedding model could understand text, images, video, audio, and PDFs all at once? Excited to share Gemini Embedding 2 our first fully multimodal embedding model.

🖼️ 5 modalities in a single unified embedding space
🌍 Supports up to 8,192 input tokens, 100+ languages
🎧 Embeds audio natively, no transcription step needed
📐 Flexible output dimensions: 3,072 / 1,536 / 768 via MRL
📎 Up to 6 images, 120s video, and 6-page PDFs per request

Now in Public Preview via Gemini API & Vertex AI.

Docs: https://lnkd.in/dutRFSqH
Blog: https://lnkd.in/d_YpkZq5

Interested in how open models works? Read this visual guide to Gemma 4 that explains all works with diagrams.

Covers how the model handles images, audio, and text, why only a fraction of parameters run during inference (MoE), and how tiny models (2B) fit on a phone using a clever embedding trick.

Worth the read even if you never plan to fine-tune anything, just to understand what's actually inside these models.

👉 https://lnkd.in/ds4zmhHP

Big Kudos to Maarten Grootendorst! LFG! 🚀

What comes after Transformers? Neural Memory and Test-Time Training. Google Research presented 2 new papers during NeurIps with an architecture that actively learns and update their own parameters during inference, acting as a "long-term neural memory" rather than a static context window.

Implementation:
1️⃣ Titans replaces the fixed-state memory of linear RNNs with a deep Multi-Layer Perceptron (MLP) memory module.
2️⃣ The model updates this memory at test-time by calculating a "surprise metric" based on the gradient of the input data.
3️⃣ MIRAS framework generalizes this by treating memory as an optimization problem with customizable loss functions and regularization.
4️⃣ Training is parallelized by chunking sequences, using linear operations within chunks and non-linear updates across chunks.
5️⃣ Models incorporate "Persistent Memory" (fixed learnable weights) alongside the dynamic "Contextual Memory" to store task-specific knowledge.

Insights:
- 💡 Attention mechanisms excellent at short-term memory but fails at efficient long-term storage due to quadratic costs.
- 📈 Deep memory structures (MLPs) significantly outperform vector/matrix-based compression used in Mamba and other linear RNNs.
- 🛠️ Memory updates are effective when driven by "surprise", high gradients indicate unexpected, memorable data.
- 📚 Forgetting mechanisms in recurrent models are mathematically equivalent to retention regularization (weight decay).
- 📉 Standard L2 (Mean Squared Error) objectives make memory sensitive to outliers; L1 or Huber loss provides better stability.
- 🧠 Titans outperforms GPT-4 on "Needle in a Haystack" tasks with 2M+ token contexts despite having fewer parameters.
- ⚡ Deep memory modules exhibit a trade-off where increased depth improves perplexity but slightly reduces training throughput.

Titans and MIRAS show potential to replace, or at least augment, pure Transformer architectures. The hybrid approach (using attention for the immediate context and Neural Memory for the deep history) suggests the future might be a convergence of RNN efficiency and Transformer performance.

Blog: https://lnkd.in/eXEEwd_t
Titan Paper: https://lnkd.in/ejxAWBJD
Miras Paper: https://lnkd.in/e27DWiyr

"LLMs are ghosts of internet-scale code" is the recent analogy from Andrej Karpathy on the Dwarkesh Patel podcast perfectly captures the current state of AI in coding.

What does this means for us as developers:

👻 LLMs are brilliant at re-creating what's common. They can instantly generate boilerplate code, scaffold a standard CRUD app, or write a routine unit test. They are channeling the "ghost" of millions of existing codebases. We should absolutely leverage this for speed. This is a massive productivity boost for repetitive tasks.

🧠 They get stuck on what's new. Ask an LLM to design a novel algorithm, implement a custom architecture, or work within a unique, non-standard codebase, and the "ghost" gets confused. It will often try to "correct" your innovative solution by forcing a familiar pattern it already knows.

AI doesn't replace fundamentals yet. it magnifies them at the moment. ✨

Deep knowledge of system design, architecture, and first-principles thinking makes you 10x more productive.

For now keep Learning how to code, keep building!

Super excited for this! 🚀 We’re rolling out a Gemini-powered personal health coach inside Fitbit (now part of Google). It uses a deep-agent architecture to orchestrate between conversational, data science, and domain expert sub-agents.

📊 Performs complex numerical reasoning on physiological time series data.
📱 Available to eligible U.S. Android Fitbit Premium users, with iOS expanding soon.
✅ Validated via 1 million+ human annotations and 100k+ hours of evaluation.
💬 Personalized guidance via a 5-10 minute interactive text or voice conversation.
🎯 Adaptive plans based on individual health metrics and goals.
🔬 Grounded in behavioral science and Consumer Health Advisory Panel.

Learn more: https://lnkd.in/eafGYwV8

We just launched Gemini 3.1 Flash Live! Our fastest, most natural real-time voice AI model for building Agents.

- Scores 90.8% on ComplexFuncBench Audio for tool use.
- 70 languages, Video streaming, Audio transcriptions, 128k context
- Comes with Agent Skill for building live voice agents.
- All generated audio is watermarked with SynthID.

Blog: https://lnkd.in/de-j3xCT
Skill: https://lnkd.in/dtdKiuRx
Docs: https://lnkd.in/d9Wu8PjA

New massive update Google TPUs Inference for open Models, like Gemma! `tpu-inference` is a new vLLM backend that delivers up to 5x performance over previous prototypes.

- 🤝 Unifies PyTorch and JAX under a single JAX-to-XLA lowering path.
- 🚀 2x-5x higher performance than the February 2025 prototype.
- 📈 20% higher throughput without model code changes.
- ⚡ Ragged Paged Attention V3 increases throughput by ~10% on Trillium (v6e).
- 🛠️ Single Program, Multi-Data (SPMD) as the default
- ☁️ Support for Trillium (v6e) and v5e TPU

Learn More: https://lnkd.in/d6mCNaQf

5 practical tips for Context Engineering, which apply to Google DeepMind Gemini as well:

1️⃣ Context Ordering Matters: Try to use "append-only" context, adding new information to the end. This maximizes cache hits reducing cost (4x) and latency.
2️⃣ Manage Tools Statically: Avoid changing tool order or availability mid-task, if not explicitly needed. This might break context caching and can/will confuse models if used tools in the history are no longer defined.
3️⃣ Use External Memory: Write explicitly or implicitly context/goals to external storage to. Preventing information loss. A typical task in Manus requires around 50 tool calls on average.
4️⃣ Recite Goals to not get lost: Prevent the model from "getting lost" by having it periodically restate its objectives. This keeps the primary goal in its recent attention span.
5️⃣ Embrace Errors: Keep error messages in the context. This allows the model to learn from its mistakes and avoid repeating them.

Learn from it: https://lnkd.in/e6n5ej7t

Excited to share a System Instructions for Gemini 3 Pro that improved performance on several agentic benchmarks by around 5%. We collaborated with Google DeepMind post-training research team to include some tips our docs. 🚀

To maximize reliability in multi-step workflows, you should craft instructions that explicitly control how the model reasons and plans. While Gemini provides strong general reasoning, complex agents benefit from prompts that enforce specific behaviors like persistence in the face of issues, risk assessment, and proactive planning.

SI: https://lnkd.in/eFYG_GsK
Docs: https://lnkd.in/eT7gFr-m

18,000 tokens/sec even for Llama 3.1 8B is ridiculous. Even a "dumb" model as llama 3.1 would be incredibly useful at this speed.

It works by merging storage and compute, permanently etching the model parameters directly into the physical transistors of the chip.

Demo link: https://chatjimmy.ai/

Today, we're launching the File Search Tool, a fully managed RAG system built directly into the Gemini API that abstracts away the retrieval pipeline so you can focus on building. File Search provides a simple, integrated and scalable way to ground Gemini with your data, delivering responses that are more accurate, relevant and verifiable.

To make File Search simple and affordable for all developers, we’re making storage and embedding generation at query time free of charge. You only pay for creating embeddings when you first index your files, at a fixed rate of $0.15 per 1 million tokens (or whatever the applicable embedding model cost is, in this case gemini-embedding-001). This new billing paradigm makes the File Search Tool both significantly easier and very cost-effective to build and scale with.

Here are the highlights:
🆓 $0.15/m tokens for indexing, free storage and embedding generation at query time.
💰 Pay a fixed rate of $0.15 per 1M tokens only for initial indexing.
🔍 Leverage advanced vector search via the Gemini Embedding model.
🧾 Receive automatic citations to verify source document usage.
📂 Ground models with PDFs, DOCX, TXT, JSON, and code formats.
⚡ Combine parallel query results in under 2 seconds.
🛠️ Integrate directly into the existing generateContent API.

Get started with the docs: https://lnkd.in/eqhdzH79

🗃️ Context Caching Update! The Gemini API implicit caching now with 90% cost savings when your requests hit the cache! 
This means if you send a request to Gemini models with a common prefix as one of previous requests, it might be cached. No code changes needed.

Price details (Normal Input/ Cached Input) per 1M token:
- Gemini 2.5 Pro: $1.25 / $0.125
- Gemini 2.5 Flash: $0.30 / $0.03
- Gemini 2.5 Flash-Lite: $0.10 / $0.01

All prices: https://lnkd.in/eZNzzBny

Enjoy :)

Nano Banana Pro (Gemini 3 Pro Image) is here 🍌 It “thinks" through a prompt and can retrieve real-time data, such as weather forecasts or stock charts, using Google Search grounding before generating high-fidelity images starting at $0.13 per image.

- $0.134 per 2k image / $0.24 per 4K image.
- Support 10 different aspect ratio and upscaling to 2K and 4K.
- Integrated with Google Search to access real-time data.
- Allows up to 14 reference input images for compositing.
- Highly accurate and clear text rendering.
- Now available in Google AI Studio and Gemini API.

Testing this felt magical. My creativity is more limiting than the models capabilities. The potential impact is massive.

Blog: https://lnkd.in/epa2BEKp 
Dev Guide: https://lnkd.in/efhAnhMP
Docs: https://lnkd.in/e5tPsVCa

Image Prompt (with Google Search Grounding): Generate an infographic of the pizza per capita.

New Guide! ✨Learn how to deploy n8n on Google Cloud Run and build your first AI agent with Gemini 2.5.

☁️ Serverless Deployment: Run open-source n8n on Google Cloud Run for auto-scaling and reduced operational overhead.
🧠 Gemini 2.5 Integration: Native support for Google's latest models to power advanced AI agents.
🔌 MCP Support: Connect your LLMs to external tools and data sources instantly using the Model Context Protocol standard.
💾 Robust Persistence: detailed setup for Cloud SQL (PostgreSQL) ensuring your workflows never lose data, using cost-effective micro tiers.
📚 Community Power: Leverage over 600+ pre-built Gemini automation workflows to get started immediately.

Implementation:
1. Install the Google Cloud CLI and authenticate with your GCP account.
2. Create a new GCP project, enable billing, and activate necessary APIs for Cloud Run, SQL, and Secret Manager.
3. Provision a Cloud SQL PostgreSQL instance for persistent storage.
4. Secure database credentials and encryption keys using Google Secret Manager.
5. Create a service account with permissions to access Cloud SQL and the stored secrets.
6. Deploy the n8n container to Cloud Run, ensuring CPU throttling is disabled to support background workflows.
7. Access the deployed n8n instance and initialize the owner account.
8. Configure Google Gemini API credentials within n8n.
9. Build a new workflow using the "AI Agent" node connected to the Gemini Chat Model.
10. Extend the agent's capabilities by adding tools, such as an MCP Client for external data access.

Blog: https://lnkd.in/eeJjUagN

Here is how Spotify uses coding background agents for thousands of code migrations and what they learned:

- Define desired verifiable end states explicitly, not strict todo steps.
- Code examples improved output reliability.
- Agent has access to 3 tools, verify, git and bash.
- “verify” tool runs formatters, linters, and tests > AGENTS md.

https://lnkd.in/dsth8x4w

We created a GitHub repo for all MCP at Google.

Get info on our remote managed MCP servers, open source MCP servers, examples, and learning resources.

github.com/google/mcp

Excited to introduce WeatherNext 2 🌦️ A new AI model from Google DeepMind and Google Research delivering faster, higher-resolution global weather predictions.

- Generates forecasts 8x faster, requiring under one minute on a single TPU.
- Surpasses prior models on 99.9% of variables across 0-15 day lead times.
- Simulates hundreds of possible weather scenarios from a single data input.
- Delivers high-resolution global forecasts down to specific 1-hour intervals.

Learn More: https://lnkd.in/e8NvFUpj

Should we add this to the Gemini API and Google AI Studio? 🤔

Nano Banana 2 is here! Pro-level image generation and editing at Flash speed with Images from web search.

💰 $0.045 per 512x image, $0.067 per 1024x image.
🌍 Improved image quality, consistency & i18n text rendering.
🖊️ Supports Text -> Text + Image(s) and Image + Text -> Text + Image(s).
🖼️ Default 1024x output, with new 512x, 2048 and 4096 output resolution.
🔍 New Text + Image search to inform generation with both text and images.
📐 Aspect ratios: 1:1, 1:4, 4:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9, 1:8, 8:1.

Available in Gemini API, AI Studio, Gemini App, Antigravity and Vertex !

https://lnkd.in/d9a_-KfU

If 2025 was the beginning of agents, 2026 will be around Agent Harnesses. An Agent Harness is the infrastructure that wraps around an AI model to manage long-running tasks. It is not the agent itself. It operates at a higher level than agent frameworks. The harness provides prompt presets, opinionated handling for tool calls (Human in the loop), lifecycle hooks or ready-to-use capabilities like planning, filesystem access or sub-agent management.

As Benchmarks are going to become more complex we need to bridge the gap between benchmark claims and user experience. A Agent Harness can be essential for three critical reasons:
- Validating Real-World Progress: Allows users to easily test and compare how the latest models perform against their use cases and constraints.
- Empowering User Experience: Without a harness, the user's experience might be behind the model's potential.
- Hill Climbing via Real-World Feedback: A shared, stable environment (Harness) creates a feedback loop where researchers can iterate and improve ("hill climb") based on actual user adoption.

We are heading toward a convergence of training and inference environments. We see a new bottleneck being context durability. The Harness will become the primary tool for solving "model drift”.

Read more in my blog: https://lnkd.in/dzW2tmkh

Learn how to build your own AI agent from scratch with Gemini 3 Pro. Excited to share this practical guide designed for everyone starting from simple text generation to a functioning CLI Agent.

🛠️ Construct a working prototype in under 100 lines of code.
🚀 Start from basic text generation using the Gemini 3 Pro model.
⚙️ Define function capabilities using JSON schemas.
🔄 Build a logic loop to intercept and execute tool calls.
💻 Assemble components into a functional multi-turn CLI application.
📚 Learn Best Practices for Engineering Agents

Blog: https://lnkd.in/ewKj9uzp

Very proud and excited to share this! 🤗

An `AGENTS.md` (or equivilant) has the highest configuration point for agents. It's injected into every conversation. But research shows that doing it wrong actively hurts performance. Here's how to do it right, backed by data

## Less Is More

- Auto-generated AGENTS.md reduce success rates by ~3% while increasing inference cost by over 20%
- Human-written AGENTS.md only marginally improve performance (~4%)
- Stronger models don't generate better context files.
- Codebase overviews in AGENTS.md don't help agents navigate faster.
- LLM-generated files are redundant with existing docs.
- Instructions ARE followed. Agents respect AGENTS.md instructions, but unnecessary requirements make tasks harder.

## What to Include

- The WHAT: Your tech stack, project structure, and what each part does. Critical for monorepos.
- The WHY: The purpose of the project and its key components. Help the agent understand intent, not just structure.
- The HOW: How to build, test, and verify changes. Include non-obvious tooling (e.g., `uv` instead of `pip`, `bun` instead of `npm`). Tools mentioned in AGENTS.md get used 160x more often than unmentioned ones.

## What NOT to Include

- Detailed codebase overviews or directory listings. The paper found these don't help agents navigate faster, and agents can discover structure themselves.
- Code style guidelines. Use linters and formatters instead, they're faster, cheaper, and deterministic.
- Task-specific instructions that only apply sometimes. Since AGENTS.md goes into every session, non-universal instructions dilute focus.
- Auto-generated content. Don't let the agent write its own AGENTS.md. The data shows this hurts more than it helps.

## How to Structure It

- Keep it short. General consensus is <300 lines; HumanLayer keeps theirs under 60 lines. Every line goes into every session, make each one count.
- Use progressive disclosure. Don't put everything in AGENTS.md. Instead, keep task-specific docs in separate files (e.g., `agent_docs/running_tests.md`, `agent_docs/database_schema.md`) and list them in AGENTS.md with brief descriptions so the agent reads them only when relevant.
- Prefer pointers over copies. Reference `file:line` locations rather than embedding code snippets that will go stale.
- Write it yourself, deliberately. A bad line in AGENTS.md cascades into bad plans, bad code, and bad results across every session.

Paper: https://lnkd.in/dh7wqU2z
Blog: https://lnkd.in/d6ih8DuM).

Gemini 3.1 Flash-Lite is here! 🔦 Our fastest, most cost-efficient gemini model built for high-volume workloads at scale.

💰 Priced at $0.25/M input, $1.50/M output tokens
🧠 Matches 2.5 Flash quality at Flash-Lite cost
⚡2.5x TFT and 45% faster output vs 2.5 Flash
💽 Enables low-latency entity extraction, classification or data processing

Now in preview on AI Studio, Gemini API & Vertex AI. 🔽

Try in AIS now: https://lnkd.in/d-n7hPDn
Blog: https://lnkd.in/dg6zAy3b

I have been using Gemini 3 Pro for a bit here are my best Practices for General Usage. Including principles and structural patterns that are currently working best for me.

This isn't meant to be treated as the gold standard, but rather as a starting point to help you refine your own strategies. Take what works, tweak what doesn't, and keep iterating.

https://lnkd.in/eAa9gFD9

Introducing the Gemini Docs MCP Server, a local STDIO server for searching and retrieving Google Gemini API documentation. This should help you build with latest SDKs and model versions. 🚀

- Run the server directly via uvx without explicit installation.
- Performs full-text search across all Gemini documentation pages locally.
- Passed 114/117 for Python and Typescript using latest SDKs and Model.
- 3 Tools: search_documentation, get_capability_page, get_current_model.
- Utilizes a local SQLite database with FTS5 for efficient querying.
- Works with VS Code, Cursor, Gemini CLI … and every other tool support MCP.

Github Repository: https://lnkd.in/eTyzmAR2

Note: Thats currently a side project of me. Should we make an official one?

More Gemma! Meet TranslateGemma, a new collection of open translation models built on Gemma 3 designed for high-performance communication.

- Available in 4B, 12B, and 27B parameter sizes.
- Evaluated on 55 languages using the WMT24++ dataset.
- 12B model outperforms the Gemma 3 27B baseline.
- 4B model optimized specifically for mobile and edge deployment.

Models: https://lnkd.in/d2KGtZiq
Technical Report: https://lnkd.in/dXMnHGrH