Open LLM Deploy

DataPrep-Bench introduces the first unified evaluation for LLMs preparing training data, covering construction from raw sources and quality scoring...

No significant updates today.

No significant updates today.

OpenForgeRL tackles a key open-source gap: complex agent harnesses like OpenClaw power real deployments but resist end-to-end RL training with...

Self-hosted LLM apps face unique hurdles like chained agentic calls, non-deterministic outputs, and unpredictable user intents that break traditional...

New Model Releases

• 🔥 Instella-MoE: AMD released Instella-MoE, a fully open 16B-parameter Mixture-of-Experts model with 2.8B active parameters,...

Kimi K3's architecture diagram has been reconstructed and released publicly, introducing KDA and AttenRes components on a K2.5 baseline. Model weights remain unavailable.

MCP servers and AMD's Venice-Helios rack-scale systems are shifting enterprise AI from pilots to production autonomous workflows by enabling secure,...

Nvidia, Microsoft, Meta, and others signed a letter urging lawmakers to skip "premature restrictions" on open-weight models that could stifle...

• Instella-MoE is a fully open 16B-total / 2.8B-active MoE model with competitive benchmark results against dense and MoE baselines.
• AMD releases...

Proxy metrics like parameter count or FLOPs approximate cost but fail to predict precision in lightweight LLMs. Direct PTME measurements (precision,...

New Open-Source Agents

• 🔥 OpenWorker: OpenWorker is an open-source agent that delivers finished work like documents or calendar updates,...

Three new frameworks highlight the push for agents that deliver real work without heavy scaffolding.

• NVIDIA's NOOA treats agents as native Python...

New benchmarks target real work while fighting contamination and measuring end-to-end performance.

• Task breakdowns reveal popular coding benchmarks...

PrismGuard offers a practical, auditable firewall for self-hosted LLMs that logs every decision with policy, threshold, and resolution details.

-...

Harness Handbook creates a three-level map linking runtime behaviors to source code via static analysis and LLM structuring.

• BGPD workflow steers...

AMD's Ryzen AI Software 1.8 adds support for more LLMs, embeddings, text-to-speech, and Stable Diffusion models, plus optimizations for running larger...

New Releases

• 🔥 Lemonade 11.5 Local AI Server: Lemonade 11.5 was released today with a completed router for routing queries to models based on...

When leading models sit within a few benchmark points of each other, the headline numbers prove far less useful than the differences behind them, according to a comparison of 33 models across 15 providers.

• Lemonade 11.5 adds completed Router for policy-based or LLM-driven query routing to models on AMD NPUs, GPUs, and CPUs.
• Hermes AI Agent now runs...