TokenPost topic
Open-weight models
11 TokenPost stories are tagged Open-weight models. The newest appear first.
Latest Open-weight models stories
- Perplexity Begins Private Preview of Orchestrator API for Developers
The catalog includes models from Moonshot AI, Z.AI and DeepSeek. Developers must request preview access by email and use an existing Perplexity API key.
· Technology
- Kimi K3.1 Pro (Max) Ranks First Among Open-Weight Agent Models
It placed 13th across all models, with results drawn from more than 17,500 extended agent-task sessions submitted by users globally.
· Technology
- Ricursive Says Ripple-65B Preview Scored 80.52 on Coding Benchmark
The 65-billion-parameter open-weight model supports context windows of about 1 million tokens. Ricursive put research and development costs at about $32 million.
· Technology
- Liquid AI Debuts Six Open-Weight Models for Local AI Workloads
The family covers text, vision-language, audio and decision tasks, alongside an omni model that handles text, images, video and audio.
· Technology
- Nvidia Releases Open-Weight Bolt Model for Local AI Agents
The model runs in 24GB of memory. Nvidia also released a 120-billion-parameter enterprise model and plans a smaller Bolt version later in October.
· Technology
- Reflection AI Releases Beam Weights With Post-Training Workflows
Beam has 501.3 billion total parameters, supports a one-million-token context window and includes a separate speculative-decoding drafter.
· Technology
- Meta’s Muse Spark Open Scores 39, Leads U.S. Open-Weight Models
It beat NVIDIA’s Nemotron 3.5 Super VL by 13 points but trailed Chinese models globally. Agent tests were its weakest category.
· Technology
- OpenRouter Users Sent 84% of Tokens to Non-Frontier Models in Aug. 10 Week
Six models handled about 80% of OpenRouter volume during the week of Aug. 10 while delivering roughly 77% of frontier-model performance at a fraction of the cost.
· Technology
- GLM-5.3 Nearly Matches Anthropic Model in Cyber Exploit Tests
The open-weight model produced working exploits in 50 of 410 attempts, while U.S. testing ranked it the most capable open-weight model evaluated to date.
· Technology
- Mistral Previews 1 Trillion-Parameter Large 4 AI Model
The open-weight system targets coding, cyberdefense, finance and other specialized work, with weights planned after safety testing.
· Technology
- Reflection AI Announces Model Aimed at Chinese Open-Weight Rivals
The Nvidia-backed startup has not released the model or published performance benchmarks.
· Technology