Google DeepMind recently released Gemma 4 12B, a dense multimodal model that completely strips traditional encoders. Vision and sound flow straight to the core of the LLM. The result is a model that runs an agent workflow on a consumer laptop with 16 GB of RAM. It is distributed under the Apache 2.0 license. Model … Read more
Your AI agent is good at testing. Then you send it, and something breaks. A tool called loops forever, like it never learns. The retrieval step returns garbage and costs more. You have absolutely no idea why. That’s the problem with agent visibility. And if you build with LLMs, you need to solve it before … Read more
Nous Research released it Hermes Desktop in public preview. It is a native operating system for macOS, Windows, and Linux. It provides the open source Hermes Agent with a user interface. So far, users have run Hermes through the CLI and messaging gateways. The current build is Hermes Agent v0.15.2. According to Nous Research documentation, … Read more
The NVIDIA AI team has been released Cosmos 3. It is a family of omnimodal world models for body AI. Models include physical reasoning, world generation, and action. All three capabilities live within one open model. NVIDIA has open sourced benchmarks, training documentation, deployment tools, and datasets. I Cosmos 3 targeted rollout of robots, autonomous … Read more
In this lesson, we are fine tuning Liquid AI’s LFM2 model by using a complete open source workflow. We start by loading the basic LFM2 testbed with QLoRA, prepare a dialog-style supervised fine-tuning data set, train a lightweight LoRA adapter using TRL and PEFT, and integrate the adapter back into the model. We also extend … Read more
If you’ve ever tried to deploy an AI agent in production, you know that the hard part is usually not the model. It’s all around: sandboxing, state administration, authentication management, tool execution, error detectionand all the infrastructure that turns a prototype into something reliable. Anthropic’s Claude Managed Agents makes that easy by providing you with … Read more
Alibaba’s Qwen team has released Qwen3.7-Plus. The model is now available through Bailian’s Alibaba Cloud platform. Bailian is a console that international users access as Model Studio. It provides API services to external developers. The release follows Alibaba’s launch in May of the Qwen3.7 generation. Qwen3.7-Plus Qwen3.7-Plus is a large multi-language model. The model understands … Read more
print(“n### SECTION D: end-to-end Transformer (vanilla fp32 vs Apex fused + AMP) ###”) VOCAB, D, NHEAD, LAYERS, SEQ, BATCH, STEPS = 2000, 256, 4, 4, 128, 32, 60 class Block(torch.nn.Module): def __init__(self, d, nhead, norm_cls): super().__init__() self.attn = torch.nn.MultiheadAttention(d, nhead, batch_first=True) self.ff = torch.nn.Sequential(torch.nn.Linear(d, 4 * d), torch.nn.GELU(), torch.nn.Linear(4 * d, d)) self.n1, self.n2 = … Read more
Hermes Agent already remembers from every session. An open source agent from Nous Research ships with selected memory files and full-text session search. But a new community project says the built-in memory is too shallow to handle heavy workloads. A new library called ‘Memory OS’ has been released under the MIT license by the developer … Read more
The Transformer’s focus has not changed since 2017. Most efficient work has tried to replace the softmax focus directly. The new paper takes a different route. Maintains softmax attention and bolts on the repair branch. A team of researchers from Northwestern University, Tilde Research, and the University of Washington present a Local Linear Attention called … Read more