NaiveAI's Naive-N0.5-Flash, an open-weights coding model on Xiaomi's MiMo-V2.5, reports near 1,000 tokens per second

Naive-N0.5-Flash, an open-weights coding model from Chinese company NaiveAI, reports server-side throughput approaching or passing 1,000 tokens per second with thinking enabled, according to testing published by Latent Space. Latent Space's headline says NaiveAI promises 2,100 tokens per second. Pricing is not yet disclosed.

What the model is

Naive-N0.5-Flash is tuned to generate as many tokens per second as possible for AI coding tasks. NaiveAI pairs it with its own NaiveRT inference engine. The company describes it as a model built largely by AI and trained to do AI research itself, with recursive self-improvement as the long-term goal.

The model evolves from Xiaomi's MiMo-V2.5, a 310B-parameter sparse Mixture-of-Experts model with 15B active parameters, trained on 48T tokens. MiMo-V2.5's language backbone inherits the hybrid design of MiMo-V2-Flash, which mixes sliding-window attention with global self-attention.

Through multi-stage continued training, NaiveAI adapted the checkpoint to DeepSeek Sparse Attention layers. According to the NaiveAI technical blog, this took roughly 3.25T additional tokens on top of the original 48T. About 3T of those were trained at a fixed 1M context length with a constant learning rate.

Published benchmarks

Latent Space compared the model with DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash. Artificial Analysis statistics for NaiveAI models are not yet available. The Intelligence Index scores for the comparison models are 45 for DeepSeek V4.1 Flash 1017 (Max) and 43 for GLM-5.3-Flash.

Scores below come from each vendor's own model card or technical blog, not from Latent Space's testing:

Model	NL2Repo-Bench	DeepSWE v1.1
GLM-5.3-Flash	56.3	63.4
DeepSeek V4.1 Flash 1017	68.2	77.0
Naive-N0.5-Flash	71.9	67.8

These figures place its coding ability between the two comparison models, and ahead of both on NL2Repo-Bench. On Agent's Last Exam, NaiveAI reports 32.4 and GLM-5.3-Flash 26.3.

Latent Space speed tests

Latent Space ran 30 text-only questions per model: 10 each from MATH-500, GPQA-Diamond and HLE-Diamond, with a 256K completion-token limit. Throughput figures come from different sources: OpenRouter for DeepSeek and GLM, and server-side measurements for Naive-N0.5-Flash. Latent Space therefore does not claim an exact speed multiple.

Median end-to-end response time:

• Naive-N0.5-Flash (thinking): 3.1 seconds
• GLM 5.3 Flash (high): 5.3 seconds
• DeepSeek V4.1 Flash 1017 (high): 10.1 seconds

Mean response time:

• Naive-N0.5-Flash: 45.3 seconds
• DeepSeek V4.1 Flash 1017: 45.0 seconds
• GLM 5.3 Flash: 86.9 seconds

On the hard questions, Naive-N0.5-Flash (thinking) averaged 89.8 seconds against 79.7 for DeepSeek V4.1 Flash 1017 (high).

Verbosity varies sharply. On hard HLE questions, Naive-N0.5-Flash with thinking averaged about 108,500 completion tokens, roughly nine times the GLM 5.3 Flash (high) average and 11 times the DeepSeek V4.1 Flash 1017 (high) average. Across all 30 questions, its median completion was about 3,200 tokens against a mean above 51,000, because a few very long reasoning traces pulled the average up. Latent Space says the model takes longer to reason through harder problems yet answers faster overall, at a slightly higher cost per question.

Latent Space notes two caveats: the sample is small, and it did not grade answer accuracy.

App-building prompts

Latent Space also gave six model and thinking-effort variants three prompts: an interactive 3D Earth and Solar System, a seeded 3D tree, and a browser port of Ray Tracing in One Weekend. On the Solar System prompt, Naive-N0.5-Flash (thinking) produced the best result of the three models, with sunspots on its sun and city lights on the dark side of its Earth. DeepSeek V4.1 Flash 1017 (high) came close behind.

What this means

For latency-sensitive coding workflows, the median figures are the most relevant data: 3.1 seconds against 5.3 and 10.1 for the comparison models. The means narrow that gap, and on hard questions Naive-N0.5-Flash was slower than DeepSeek V4.1 Flash 1017, so the advantage depends on the workload. The throughput comparison is also not like for like, since NaiveAI's figures are server-side and the others come from OpenRouter. Published benchmarks show a mixed picture: first on NL2Repo-Bench, behind DeepSeek on DeepSWE v1.1. Without graded accuracy or an Artificial Analysis index score, quality relative to the comparison models remains unquantified by independent testing.