The Pulse
NaiveAI Releases 309B Model With 1M-Token Context
NaiveAI announced Naive-N0.5-Flash, a 309-billion-parameter open-weight model with 15.5 billion active parameters and a native 1-million-token context window. The company says it uses AI-assisted research and a custom inference system for c
AI.info Team · Written with AI · fact-checked ·
NaiveAI’s new model has 309 billion parameters, but activates 15.5 billion for each token it processes. The company announced Naive-N0.5-Flash on September 27 as an open-weight model for coding and AI research, with a native context window of 1 million tokens.
The release’s central engineering claim is that the model handles that context without any full-attention layers. NaiveAI says its weights and inference code are released under the MIT license; its performance figures, however, come from the company’s own tests.
One million tokens without full attention
Naive-N0.5-Flash is a mixture-of-experts model, a design that activates only part of its total parameters for a given token. NaiveAI says its attention architecture combines Sliding-Window Attention with DeepSeek Sparse Attention, using a predominantly 5-to-1 ratio. The company describes the network as entirely local or sparse, with no full-attention layers.
That architecture is intended to make very long inputs practical without applying full attention across the entire sequence. The release targets coding and AI research and development, including research tasks that can accumulate long histories of observations, tool outputs and prior exchanges.
NaiveAI reports 2,122 tokens per second
NaiveAI’s research page gives two related speed figures. Its overview says the NaiveRT inference system can reach up to 2,000 tokens per second; a later benchmark section reports a peak single-stream rate of 2,122 tokens per second on eight GPUs.
The 2,122 figure is measured over the best one-second window, with thinking turned off and prefill excluded, across 41 HTML and SVG generation requests. The company also reports that a full speculative-decoding round took 3.4 milliseconds on its system, compared with 12.3 milliseconds for SGLang on the same hardware. Those are company-reported measurements, not an independent evaluation.
AI helped build the model—and its runtime
NaiveAI says AI systems helped explore and design the model’s hybrid attention architecture and optimize its training, inference and deployment systems. The company describes human researchers as setting direction and deciding which work to accept.
“Human researchers set direction, define constraints and criteria, and make critical decisions.”
NaiveAI Team
The company presents the model as both a tool for AI research and a product of AI-assisted research. NaiveAI says it built NaiveRT through 151 documented optimization trials, adopting 63 changes and rejecting or rolling back others. It says the runtime’s source code and benchmark materials will be available by October 12.
Weights are released; API access is still to come
NaiveAI says the model weights and inference code are released under MIT terms. The company also lists planned API pricing of $0.10 per million input tokens, $0.40 per million output tokens and $0.01 per million cache-read tokens, while describing API access as forthcoming.
The release leaves practical tests for users: whether the model sustains useful quality across a million-token context, and whether NaiveRT’s reported speed holds across workloads and hardware beyond the company’s benchmark setup.