Open Source AI

Tencent released Hy3 under Apache 2.0: a 295B MoE model with strong results on reasoning and agent tasks.

• Day-0 vLLM support and production kernels...

A verification loop boosted DeepSeek from ~20 to ~80 on Web-Bench, matching Opus performance at roughly 1/7th the cost. This practical result aligns...

Chinese open-weight models now handle more OpenRouter tokens weekly than all US models combined.

• Hy3 release: Tencent's 295B MoE model activates...

NanoQuant delivers the first post-training method to quantize LLMs below 1-bit via low-rank binary factorization and ADMM initialization, achieving...

John Carmack argues that deterministic access patterns in model inference eliminate the need for random-access memory, opening the door to NAND flash over 100x cheaper than HBM for specialized accelerators that could run much larger models locally.

New Open Weights Releases

• 🔥 Tencent Hy3: Tencent released Hy3, a 295B MoE model with 21B active parameters, 256K context, and Apache 2.0...

Open AI is expanding rapidly across model sizes and geographies.

• Small specialized models now power critical tools in Africa and Asia, like...

Buying guides spotlight VRAM as the key spec for running 7B-70B models locally, with options like the RTX 5060 Ti 16GB at $560 or the 32GB RTX 5090...

Ghost memory causes agents to confidently repeat outdated user facts weeks later because old, current, and transition records all sit in the same...

Government intervention forcing Anthropic to pull Fable has ended Silicon Valley's "move fast and break things" era. This regulatory pressure signals...

Hy3, Tencent's new 295B MoE model, delivers flagship-level performance in its size class under a permissive Apache 2.0 license that's explicitly...

Enterprise platforms, local tools, and cloud services each address different open source AI needs.

• Enterprise internal platforms evolve via...

Developers struggle to translate everyday tasks like bug fixes and log filtering into precise prompts or code, but a new Program-as-Weights approach...

Mini PCs outperform Edge AI boxes for real agent workloads thanks to upgradable RAM, sustaining 12+ t/s on OpenClaw tasks while 16GB unified-memory...

• Licensing breakthrough: Tencent's Hy3 and Alibaba's Qwen now ship Apache 2.0, unlocking EU/UK enterprise use previously blocked by restrictions
  -...

Adversarial fine-tuning lets attackers embed hidden malicious behaviors in LLMs that bypass standard benchmarks.

• Attackers manipulate training...

Enterprise AI teams often pick the wrong customization method by asking which technique is more advanced rather than which fits the problem.

• RAG...

The debate pits AI's promise of instant custom tools against community-built ones that carry years of real-world scars.

• Throwaways die: Simple...

New Open Releases

• 🔥 Mistral Leanstral 1.5: Mistral released Leanstral 1.5, an Apache 2.0 open-source model that solved 587 of 672 Putnam...

Open-source AI is carving out specialized roles in rigorous research domains.

• OpenScience delivers a fully Apache 2.0 workbench that runs locally,...