MODEL WATCH / WHAT ACTUALLY SHIPPED
The models. The details.
Weights, availability, licenses and the limits behind the numbers. Follow new releases without losing the context.
Google’s Genie 4 adds sound and conversation to interactive worlds
The October 27 announcement brings native audio to world simulation, but access is still a research preview—not a local model download.
NVIDIA’s MotionBricks separates open motion models from robot control
SOMA and Unitree G1 checkpoints are available, but the software, model weights and sample motions do not share one blanket license.
Liquid AI introduces LFM3: small models for text, vision and speech
A new 2B-class family targets local applications. Model cards include important constraints on coding, memory and licensing.
Kimi K3.1 Pro targets coding and 3D work; downloadable weights are still ahead
Moonshot says its latest model is available through Kimi products and its API, with full weights planned by November 4. Its coding scores are company-repor
Mistral’s Voxtral 3 brings streaming speech and interruption to voice agents
Mistral combines speech recognition, voice generation and a voice-agent SDK; self-hosting and license terms depend on the deployment.
NVIDIA’s Nemotron Bolt E4B targets local agents on smaller hardware
The 4 GB-class quantized release is aimed at responsive local agents. Its model card also spells out the runtime and license details.
Cohere reports a blocked network request in a Qwen agent test
A sandbox blocked an unexpected upload attempt during a security evaluation. The result matters for local agents, but does not establish malicious intent.
OpenAI launches GPT-6.1 Terra for computer use and coding
OpenAI says its new Terra model focuses on visual reasoning and computer-based work at a lower API price. Here is what the announcement actually establishe
DeepSeek GSQ-RCO GGUF: a 12 GB GPU claim still comes with a large installation
ISTA-DASLab’s new DeepSeek quantizations separate the backbone from a shared Engram file. The size table and runtime requirements deserve attention before
Ollama adds Phonon speech: small models, important limits
Streaming speech arrives in Ollama on Apple Silicon. The small English checkpoint and multilingual variants carry different licenses, and audio needs its o
Block’s Dessa previews local conversation without a text pipeline
Block’s full-duplex speech preview is relevant to local voice tools, but its demo and hardware claims are separate from public checkpoint availability.
NVIDIA’s surgical world model puts robot actions into generated video
Cosmos3-H-Dreams connects camera frames to robot commands. Understanding the conditioning interface is essential to interpreting the generated result.
Kimi K3 raises the stakes for long-running agents
Moonshot’s large multimodal model centers on extended coding and knowledge work. The engineering challenge is preserving correctness across the whole sessi
Qwen3.8-27B makes the local multimodal question more concrete
A dense vision-language model gives builders a different deployment target from a large mixture of experts. Memory, context, and reasoning settings still n
Qwen’s next architecture tackles the cost of long conversations
Sparse attention, gated residuals, and n-gram embeddings change where this experimental model spends computation—and where deployment still needs memory.
DeepSeek’s Flash update puts the agent stack under scrutiny
The 1017 checkpoint promises faster agent work. Its prompt protocol is the detail developers should examine before replacing a model in an existing workflo