NaiveAI/Naive-N0.5-Flash model card

Fetched 2026-10-18. HF API createdAt 2026-09-27T14:14:32Z, lastModified 2026-09-27T15:16:00Z, sha 9c9a11e8d2a6b8507d48b533345ef591cf9a6494, license MIT, usedStorage 617,823,283,433 bytes across 49 safetensor shards. Card: open-weight MoE continued from Xiaomi MiMo-V2.5, 309B total, 15.5B active, native 1M context, 48 layers of which 39 are sliding-window attention (window 128) and 9 are DeepSeek Sparse Attention selecting the top 2048 tokens, GQA with 4 KV groups, no full-attention layers. The README does not print a numeric benchmark table. The numbers live in assets/coding-benchmarks.pdf (and the matching alt text on https://naive.ai/en/research/, fetched the same day): NL2Repo-Bench 71.9, DeepSWE v1.1 67.8, Terminal-Bench 2.1 86.7, SWE-bench Pro 73.6, ProgramBench Almost@1 17.5, Agents' Last Exam CLI 32.4, FrontierSWE v1 78.2. The card says its own scores were run with Claude Code 2.1.207, a 1M context, temperature 1.0, top_p 0.95, and only file I/O plus Bash. Comparison rows on that figure (DeepSeek-V4.1-Flash NL2Repo 64.0, GLM-5.3 58.0, Kimi-K3 58.0) are the vendor's own readings of other cards, not this catalog. Deployment section: FP8 mixed precision, transformers>=5.17.0 with trust_remote_code, weights about 315 GB, custom code (configuration_naive_n05_flash.py, modeling_naive_n05_flash.py). The model tree links llama.cpp, LM Studio, Jan and Ollama quantizations (4 community repos) but the vendor card does not publish a GGUF recipe. No SWE-bench Verified, GPQA Diamond, AIME or MMLU-Pro number is printed.