Latent Space test: NaiveAI's Naive-N0.5-Flash hits about 2,100 tokens/s, 3.1s median response

Latent Space's first model evaluation measured NaiveAI's open-source Naive-N0.5-Flash at a peak of about 2,100 tokens per second and a median response time of 3.1 seconds with thinking enabled, according to a report by QbitAI (量子位), a Chinese-language outlet.

Latent Space launched a new evaluation column on November 10, described as a no-stance test of new models. Its first subject is Naive-N0.5-Flash from China's NaiveAI. The comparison models are DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash, which QbitAI describes as two of the most popular Flash models on OpenRouter.

Benchmark results

On overall capability, Latent Space places Naive-N0.5-Flash between DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash. On NL2Repo-Bench, which tests building a complete code repository from natural-language requirements, Naive-N0.5-Flash scored 71.9. That is 3.7 points ahead of DeepSeek and 15.6 points ahead of GLM, per the report.

Speed

According to NaiveAI, single-user output is about 50 tokens/s in standard mode and up to 2,000 tokens/s in fast mode. Latent Space's own test used 30 questions across math, science knowledge and hard reasoning, run on all three models. It measured:

• Peak speed of about 2,100 tokens/s
• About 984 tokens/s in the thinking phase
• 1,816 tokens/s in the answering phase
• Median response time with thinking on: 3.1 seconds for Naive-N0.5-Flash versus 10.1 seconds for DeepSeek

QbitAI states this is about 69% less time per request, or about 3.3 times faster.

Coding tasks

Latent Space gave the three models identical prompts for three web-app tasks, across different thinking modes:

1. 3D solar system: Naive-N0.5-Flash in thinking mode was judged best of the three. Sunspots were visible, and the night side of Earth showed city lights.
2. Real-time 3D growing tree: Naive-N0.5-Flash (thinking) and DeepSeek V4.1 Flash 1017 (high) both produced high-quality results and performed similarly.
3. Real-time ray tracer based on Ray Tracing in One Weekend: all three models delivered runnable applications, and NaiveAI implemented the core features.

QbitAI notes that three tests do not represent a model's full capability.

Technical approach

QbitAI says NaiveAI focuses on inference efficiency. AI participated in optimizing its in-house NaiveRT system, which pushes generation speed to a maximum of 2,000 tokens/s. The model uses sparse attention and supports a context of one million tokens. The model architecture, training system and inference optimization were all developed with AI involvement.

The company's founder is Dai Jifeng (代季峰), an associate professor in the Department of Electronic Engineering at Tsinghua University. QbitAI says NaiveAI was founded less than a year ago and has used an AI-driven R&D approach from the start: AI writes code, runs experiments and iterates, while human researchers set direction and make key decisions. Its infrastructure supports nearly ten million sandbox runs per week.

Latent Space author Gabriele Beltramo sees promise in the approach of having AI optimize its own inference CUDA kernels and look for model architectures better suited to large-scale reinforcement learning.

Pricing for Naive-N0.5-Flash was not stated in the report.

What this means

The most concrete result is latency. In Latent Space's 30-question test, the median response time with thinking on was 3.1 seconds against 10.1 seconds for DeepSeek V4.1 Flash 1017. For coding and agent workflows that chain many model calls, QbitAI argues, per-request savings accumulate into shorter task cycles. On capability, the evidence is mixed by design: the model sits between the two comparison models overall, leads on NL2Repo-Bench, and ties DeepSeek on one of the three web-app tasks. The speed figures come from a single outside test of 30 questions, and the 2,000 tokens/s fast-mode figure is NaiveAI's own claim. Pricing is the open variable for anyone weighing it against the two Flash models.