How much does the Naive-N0.5-Flash API cost?

NaiveAI lists API pricing of $0.10 per million input tokens, $0.40 per million output tokens and $0.01 per million cached input tokens as of September 27, 2026. That is cheaper on input than DeepSeek V4.1 Flash ($0.15/$0.60 off-peak), MiMo-V2.6-Flash ($0.14/$0.28) and GLM-5.3 Flash ($0.15/$0.50), and matches GPT-6 Luna's $0.10 input while costing less on output ($0.40 vs $0.50).

Can I run Naive-N0.5-Flash locally?

Only on a multi-GPU server. The weights occupy about 315 GB in FP8 and the model requires FP8-capable NVIDIA GPUs, so it is out of reach for a single consumer GPU. NaiveAI ships a Transformers quick start (transformers>=5.17.0 with the NaiveAI/Naive-N0.5-Flash-FP8 checkpoint), and its own NaiveRT inference stack claims 50 tokens/s per user in Standard mode and up to 2,000 tokens/s in Ultrafast mode.

What does 'no full-attention layers' mean for Naive-N0.5-Flash?

Most long-context models keep a few global-attention layers whose per-token decode cost grows with context length, which dominates cost at 1M tokens. Naive-N0.5-Flash replaces those with DeepSeek Sparse Attention: a lightweight indexer scores the full history and the backbone attends only to the top 2,048 selected tokens. Combined with 128-token sliding-window layers, every layer is local or sparse, so decode cost at 1M context stays close to flat. The full KV cache is still retained.

Quick Answer

What Is Naive-N0.5-Flash? NaiveAI's 309B Open MoE (2026)

Q: What is Naive-N0.5-Flash?

Naive-N0.5-Flash is an open-weight 309B-parameter Mixture-of-Experts model with 15.5B active parameters, released under the MIT license on September 27, 2026 by NaiveAI, a Beijing startup led by Tsinghua's Jifeng Dai. It is built for coding and AI R&D, supports a native 1M-token context, and uses a hybrid of 39 sliding-window attention layers and 9 DeepSeek Sparse Attention layers with no full-attention layers at all. It was continued-pretrained from Xiaomi's open MiMo-V2.5 base on 3.25T tokens.

Published: September 28, 2026

The short answer

Naive-N0.5-Flash is a 309B-parameter, 15.5B-active MoE model with a native 1M-token context and no full-attention layers, released with MIT-licensed weights on September 27, 2026 by NaiveAI. It targets coding and AI research-and-development workloads, was continued-pretrained from Xiaomi’s open MiMo-V2.5 base on 3.25T tokens, and is priced at $0.10 input / $0.40 output / $0.01 cache read per million tokens on NaiveAI’s API. The company’s claim to fame is that the research pipeline itself was “substantially executed by AI systems,” with humans setting objectives and evaluation standards.

Who is NaiveAI

NaiveAI is a Beijing startup led by Tsinghua University’s Jifeng Dai, reportedly valued at about $1.4 billion. Its tagline is “Building Frontier AI with AI”: the stated method is to have AI systems run most of the research and engineering loop — architecture search, kernel optimization, training decisions — under human-defined objectives. Naive-N0.5-Flash is its first open-weight release and the NaiveRT inference stack is presented as a case study of the same process.

Architecture in one table

Property	Naive-N0.5-Flash
Type	Mixture-of-Experts
Total / active parameters	309B / 15.5B
Base model	Xiaomi MiMo-V2.5 (open weights)
Context	Native 1M tokens
Layers	48 (eight six-layer modules)
Attention layout	39 sliding-window (128-token window) + 9 DeepSeek Sparse Attention layers; zero full-attention layers
DSA selection	Top 2,048 tokens per query, GQA with 4 KV groups, 16 indexer heads
Training	3.25T tokens: 50B indexer warm-up, 3T sparse-attention training, 200B LR decay
Precision	FP8 mixed-precision inference; ~315 GB weights
License	MIT (weights + inference code)
API price	$0.10 in / $0.40 out / $0.01 cache read per MTok
Inference	NaiveRT: 50 tok/s per user Standard, up to 2,000 tok/s Ultrafast

Why the attention design is the story

Long-context MoE models from 2025-2026 typically use mostly sliding-window layers with a handful of global-attention layers to keep long-range information. At a million tokens those few global layers dominate decode cost. NaiveAI swapped them for DeepSeek Sparse Attention (DSA), which the DeepSeek team introduced with V4: a small indexer scores the whole history, then the main attention only reads the 2,048 highest-scoring tokens. NaiveAI also replaced DeepSeek’s MLA with grouped-query attention (four KV groups), which is simpler to serve on standard stacks.

The practical result is that per-token decode cost does not climb with context length the way it does in models that keep full-attention layers. The trade-off is that the indexer still scans the full history and the full KV cache is kept in memory, so memory footprint at 1M tokens is not reduced — only compute and memory bandwidth are.

How it was evaluated

NaiveAI’s model card says the coding and agentic evaluations were run inside Claude Code 2.1.207 with a 1M-token context, temperature 1.0, top-p 0.95, with the harness exposing only basic file I/O and Bash tools. Comparison scores are pulled from vendor pages for GLM-5.3, GLM-5.3-Flash, Kimi K3, Qwen 3.8 Max, Hy4-preview, DeepSeek V4.1 Flash, Step-5-preview and MiniMax M3, plus SWE-Bench Pro numbers for GPT-5.6 Sol, Opus 5 and Opus 5.5 from OpenAI’s and Anthropic’s own posts. The AI R&D suite (SOL-ExecBench, NanoChat AutoResearch, NanoGPT SpeedRun) uses NaiveAI’s in-house harness. As of September 28, 2026 there is no independent reproduction and no Artificial Analysis Intelligence Index score, so treat the charts as vendor-reported.

Where it sits in the open flash tier

The interesting comparison is not against frontier closed models but against the other 1M-context open MoE “flash” models that shipped in September 2026: Xiaomi’s MiMo-V2.6-Flash (310B/15B, $0.14/$0.28), DeepSeek V4.1 Flash (552B, $0.15/$0.60 off-peak) and Z.ai’s GLM-5.3 Flash ($0.15/$0.50, weights not shipped). Naive-N0.5-Flash undercuts all three on input price and shares a base lineage with MiMo. Full comparison: Naive-N0.5-Flash vs MiMo-V2.6-Flash vs DeepSeek V4.1 Flash vs GLM-5.3 Flash.

How to run it

pip install "transformers[torch,kernels]>=5.17.0"

from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "NaiveAI/Naive-N0.5-Flash-FP8"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True,
                                             dtype="auto", device_map="auto")

Recommended sampling is temperature 1.0 and top-p 0.95. Budget roughly 315 GB of GPU memory for weights plus KV cache; SGLang is credited in the acknowledgments, so expect that path to be the served one. For a single-GPU alternative, see best open-weight models for a single GPU 2026.

Caveats

Vendor-reported benchmarks only, as of September 28, 2026.
trust_remote_code=True is required — audit the modeling code before running it on anything sensitive.
The API “will also be provided”; check availability before designing around the $0.10/$0.40 rate.
The “AI did the research” framing is a claim about process, not a verifiable property of the weights.

Last verified: September 28, 2026.