🚀 RELEASES & LAUNCHES
- Agnes-AI releases 33B open-weight multimodal Agnes-3.0-Flash
Agnes-AI released Agnes-3.0-Flash, a 33B open-weight multimodal model with a 262K-token context window.- Agnes-3.0-Flash targets developers seeking local, flagship-class reasoning across text, image and video tasks.
- The model supports adjustable reasoning effort, tool calling and a 262,144-token context window.
- Its 72-layer hybrid architecture uses 54 recurrent delta-rule layers and 18 global-attention layers.
- The model card reports scores of 74.20 on IFBench, 38.08 on SciCode and 85.05 on GPQA Diamond.
- Weights are available under Apache-2.0 and can be served with Transformers, vLLM or SGLang; the architecture limits context-growing KV cache to the global-attention layers.
📊 BENCHMARKS & EVALS
- Custom runtime runs DeepSeek-V4.1 Flash on A100 GPUs at up to 62 tok/s
An open-source runtime enables DeepSeek-V4.1 Flash inference on A100 GPUs.- For developers with A100 hardware, including multi-GPU servers and single-GPU setups.
- The runtime runs the 552B MoE backbone and 196B Engram memory on 8× A100 80GB GPUs.
- It reaches 52 tok/s with layer pipelining, 62 tok/s with expert parallelism, and 32–35 tok/s on one A100 with CPU experts.
- The OpenAI-compatible server supports streaming chat and completions despite A100 lacking FP8/FP4 tensor cores.
💰 BUSINESS & FUNDING
- Perplexity deploys GPT-6 Astra across communications and production systems
Perplexity uses GPT-6 Astra to handle communications, software changes, and production monitoring.- Perplexity uses Astra across its communications and software operations.
- The system can change software and monitor production systems end to end.
- Perplexity checks in much less frequently than with earlier models.
Sources 📎 Coverage: openai.com · 👁 via OpenAI News