Naive-N0.5-Flash

Naive-N0.5-Flash is an open-weight model from NaiveAI, a Beijing start-up, published on 27 September 2026 with a research page of that date and a Hugging Face repository created the same day; RuntimeWire, AI Weekly and Pandaily reported it the same day. It is aimed at coding and AI research work. It is a mixture-of-experts model with 309 billion parameters, of which 15.5 billion are active for each token, and a native context window of one million tokens. Instead of full attention it mixes sliding-window layers with DeepSeek-style sparse attention layers (39 and 9 of its 48 layers), which keeps the cost of each new token low on long inputs. It was built on Xiaomi’s open MiMo-V2.5 base model with further pre-training, and NaiveAI says AI systems did much of the architecture search and engineering. Weights and inference code are released under the MIT licence. NaiveAI’s page announces an API at $0.10 per million input tokens, $0.40 per million output tokens and $0.01 for cached input, and says access “will be provided”; RuntimeWire quoted yuan prices (¥0.60, ¥2.60 and ¥0.07) that do not match those dollar figures, so a buyer should confirm the rate on NaiveAI’s platform before relying on it. NaiveAI reports 67.8 on DeepSWE v1.1, 73.6 on SWE-bench Pro, 86.7 on Terminal-Bench 2.1, 32.4 on Agents’ Last Exam (ALE-CLI) and 73.7 per cent on MLE-bench-30, run in its own harness, and up to 2,000 tokens per second in an “Ultrafast” serving mode (50 per user in standard mode). Those are NaiveAI’s own measurements and are not stored as scores; no independent evaluation had been published when this was read on 28 September 2026.