local-inference

Ollama v0.40.5 adds NVIDIA Nemotron 3.5 Super, a 120B-parameter hybrid Mamba-2/Transformer MoE vision-language model with 12B active parameters and context up to 1M tokens. It also adds a smaller sibling, Nemotron 3.5 Bolt, and fixes a Responses API bug with custom tool calls.

Google DeepMind has released Gemma 4 12B, a 12-billion parameter multimodal model that runs locally on laptops with 16GB of RAM. The model eliminates separate vision and audio encoders, processing raw inputs directly through its language model backbone under an Apache 2.0 license.

H Company released Holo3.1, a computer-use agent model family ranging from 0.8B to 35B parameters. The 35B-A3B variant scores 79.3% on AndroidWorld, up from 67% in Holo3. For the first time, H Company ships quantized checkpoints (FP8, Q4 GGUF, NVFP4) enabling local inference with 1.74× throughput gains and sub-4-second agent step times.

Meta's recently acquired AI startup Manus launched a desktop application enabling its AI agent to directly access local files, tools, and applications on personal computers through a 'My Computer' feature. Previously cloud-only, the move positions Manus to compete with OpenClaw, the open-source AI agent that sparked recent industry momentum. Unlike OpenClaw's free, MIT-licensed offering, Manus operates as a paid subscription service.