NaiveAI has released Naive-N0.5-Flash, a 309-billion-parameter mixture-of-experts model aimed at coding and AI research work, as open weights under the MIT license on Hugging Face, AI Weekly reported Sunday. The model supports a native context of 1 million tokens without a single full-attention layer, though the benchmark and speed figures behind the launch are self-reported.

What Happened

According to the model card, Naive-N0.5-Flash activates 15.5 billion parameters per token. Its 48 transformer layers are arranged in eight six-layer modules: 39 layers use sliding-window attention with a 128-token window, and nine use DeepSeek Sparse Attention, which selects the top 2,048 tokens to attend to. The model uses grouped-query attention with four key-value groups.

Training started from Xiaomi’s open-weight MiMo-V2.5 base and ran to 3.25 trillion tokens at the full 1-million-token context, split into a 50-billion-token warmup for the sparse-attention indexer, 3 trillion tokens of sparse-attention training and a 200-billion-token learning-rate decay phase.

RelatedAnthropic Prices Claude Haiku 5.5 at a Tenth of Its Predecessor for Most API Requests

Running the model requires FP8-capable Nvidia GPUs, and the weights alone take about 315GB. NaiveAI also offers an API priced at $0.10 per million input tokens, $0.40 per million output tokens and $0.01 per million cached reads.

The company reports a score of 73.6 on SWE-bench Pro, and its model card charts results on coding, agentic and AI-research tests including Terminal-Bench, MLE-bench-30 and PaperBench against models such as DeepSeek-V4.1-Flash, GLM-5.3 and Claude Opus variants. AI Weekly said no independent verification of those numbers had been published.

Why It Matters

Standard attention compares every token with every other token, so its compute and memory costs climb steeply as inputs get longer. The absence of any full-attention layer is the feature AI Weekly singled out: NaiveAI is betting that short local windows plus a sparse selection of relevant tokens are enough, which would make million-token sessions, such as loading a large codebase, much cheaper to serve.

The pricing reflects that bet. Coding agents tend to send very large prompts, so input cost dominates their bills, and NaiveAI has set its input price at 10 cents per million tokens.

The speed claims are where caution matters most. NaiveAI says its NaiveRT inference stack reaches up to about 2,000 tokens per second in an "Ultrafast" mode and about 50 tokens per second per user in standard mode. AI Weekly reported that NaiveRT was built by AI in six days across 151 optimization trials and hit 2,122 tokens per second on eight GPUs in single-stream decoding, but noted that those were narrow decode-only measurements taken over one-second windows.

Caveats

AI Weekly said CellCog has flagged an unresolved intellectual-property dispute between NaiveAI and MiroMind over technology claims underlying the release. CellCog was the only outlet to report that concern, and neither company’s position was detailed.

The model card credits DeepSeek for the sparse-attention design and SGLang for inference work, underscoring how much of the release builds on other groups’ open research and code.

Context

Open-weight models are taking a growing share of real traffic. Vercel said in its September AI Gateway index that open-weight models handled 56% of tokens on its gateway in August, the first month they carried the majority of volume. Releases like this one, stacked on another lab’s open base and priced near the floor, are part of what is pushing that share up.

What to Watch

Independent runs of SWE-bench Pro and the other coding tests will show whether the self-reported scores hold. Community quantizations will determine whether the model can run on hardware short of a full FP8 GPU node, and any public response from NaiveAI or MiroMind will clarify the IP dispute.