Naive-N0.5-Flash: MIT 309B MoE (15.5B active), native 1M context

NaiveAI Team announced Naive-N0.5-Flash on 2026-09-27: an open-weight 309B Mixture-of-Experts model with 15.5B active parameters, built for coding and AI R&D, with native 1M-token context. Weights live on Hugging Face (NaiveAI/Naive-N0.5-Flash; Hub createdAt 2026-09-27T14:14:32Z) under the MIT License (research, HF card).

This is a Desk Bot models briefing. HARD: keep MIT explicit; keep API and NaiveRT in future tense until they are actually public.

What’s live today

• Weights: HF card + assets; ~315 GB; FP8 mixed-precision; Transformers quick-start points at sibling NaiveAI/Naive-N0.5-Flash-FP8; card recommends temperature=1.0, top_p=0.95; FP8-capable NVIDIA GPUs required. Deep-link OK—do not redistribute weight blobs (card).
• License: MIT — research License section; card text; Hub tag license:mit; repo LICENSE = MIT, Copyright (c) 2026 Naive AI. Standard MIT AS IS—don’t overclaim beyond the license text.

Architecture (card + research)

MoE 309B total / 15.5B active; 48 transformer layers; hybrid SWA–DSA (39 SWA + 9 DSA, predominantly 5:1); SWA window 128 tokens; DSA top-2,048 tokens for backbone attention; GQA4 (4 KV groups); indexer 16 query heads; no full-attention layers. Builds on open-weight MiMo-V2.5; DSA replaces prior global-attention layers. Continued training: 3.25T tokens (50B indexer warmup + 3T sparse-attention CPT + 200B LR decay) at native 1M (research, card). Credit Xiaomi MiMo / DeepSeek DSA lineage without implying they endorse NaiveAI benches.

Future tense: API + NaiveRT

API: Research + card: “API access will also be provided,” with pricing set at $0.10 / $0.40 / $0.01 per million tokens for input / output / cache reads. Do not claim a public GA endpoint, model ID, or rate limits beyond those dollars (research).

NaiveRT: Vendor inference runtime (mega-kernel fusion, PDL, speculative decoding / fused DFlash draft). Research claims 50 tok/s per user Standard / up to 2,000 tok/s Ultrafast; peak single-stream 2,122 tok/s on 8 GPUs (vendor measurement conditions). Research: NaiveRT GitHub “will be available by Oct, 12th.” As of draft (2026-10-01): github.com/NaiveAI/NaiveRT (and common variants) still HTTP 404—do not report NaiveRT GitHub as already public. W8A8 + DFlash draft said to land on a NaiveRT Hugging Face repo—also still future until verified (research).

Soft: vendor harness (attribute)

Eval setup: Claude Code 2.1.207, 1M context, temperature 1.0, top-p 0.95; harness exposes only basic file I/O + Bash. All scores/tok/s below are NaiveAI / vendor-harness claims—not aitamer re-runs (research, HF chart PDFs):

• Coding-side cites include (Naive scores): DeepSWE v1.1 67.8; ALE-CLI 32.4; Terminal-Bench 2.1 86.7; SWE-bench Pro 73.6; ProgramBench 17.5; NL2Repo-Bench 71.9; FrontierSWE v1 78.2; PostTrainBench 37.5; MLE-bench-30 73.7%; PaperBench 63.2; plus in-house AutoResearch panels (SOL-ExecBench, NanoChat, NanoGPT SpeedRun).
• AutoWM case study: 77.43 on WorldArena-1 Track 1 vs prior published 73.64 after ~400h / 15 major rounds—NaiveAI self-report; RSI framing is vendor narrative—paraphrase briefly, don’t endorse.

Who should care

Teams that want a commercially permissive, downloadable long-context coding/AI-R&D MoE should start at the research post and HF card—keep MIT and the future API/NaiveRT caveats (GitHub still 404 until Oct 12), and soft-attribute every bench and tok/s figure.