Naive AI Open-Sources First Model Naive-N0.5-Flash: 309B Total Parameters, Native 1M Context, MIT Weights
According to a September 28, 2026 report by Machine Heart, Dai Jifeng's team at Naive AI released its first model, Naive-N0.5-Flash, with 309B total parameters, 15.5B activated, native support for a one-million-token context, and a focus on coding and AI R&D. The report says the weights and inference code use the MIT license and that an API is planned, and it also discloses evaluation results including PaperBench and MLE-bench-30.
What changed
According to a September 28, 2026 report by Machine Heart, Naive AI, founded by Dai Jifeng, released and open-sourced its first model, Naive-N0.5-Flash, with 309B total parameters and 15.5B activated parameters, native support for a one-million-token context, and a positioning aimed at coding and AI R&D. The report states that the model starts from an open-weight MiMo-V2.5 base, replaces some global attention layers with DeepSeek sparse attention to form a hybrid structure of five sliding-window attention layers plus one sparse attention layer, and completes multi-stage training with a cumulative 3.25T tokens under a 1M context configuration. The reported openness plan is that the model weights and inference code use the MIT license and that an API is planned; the report also gives per-million-token prices for input, output, and cache reads. On evaluation, the report says the model scores 63.2 on PaperBench, 73.7% on MLE-bench-30, 71.9 on NL2Repo, and 32.4 on ALE-CLI, and gives two case studies, the NaiveRT inference runtime and the AutoWM world-model research effort. These figures are results relayed from a technical report by the source and have not been independently verified.
What it means for developers
There are three layers of direct implications for developers. First, if the MIT license arrangement holds, the weights can be downloaded, self-hosted, and modified for commercial use, which differs from closed-model routes that offer only a hosted API; note, however, that weight availability is not the same as unrestricted licensing, and the actual license files and third-party component terms shipped with the repository must be checked. Second, 309B total parameters with 15.5B activated is a relatively large sparse-expert scale; even with low activated parameters, full weight loading and VRAM footprint remain far higher than for a dense model of the same activated size, so local deployment should be assessed against total, not activated, parameters. Third, with native 1M context plus hybrid sparse attention, the sparse layers still require a full KV cache and the indexer still scans the full history, so long context is not zero-cost, and whether an inference framework supports this hybrid attention structure will directly determine whether the official inference code can be reused. The reported evaluations cover paper reproduction, machine-learning engineering, and fixed-budget post-training, pointing to having the model carry out experiment implementation and iteration rather than general conversation, so selection should be evaluated against that positioning.
Integration and evaluation
Two actionable decision suggestions, both conditional analysis rather than reported facts. First, on deployment form: the report says the weights and inference code are MIT-licensed and that an API is planned, so developers need to decide first between self-hosting and using the API. If the workload includes large amounts of experiment logs, code, and tool return values that should not be sent externally, self-hosting is more suitable, provided there is VRAM or a quantization scheme able to accommodate 309B total parameters; if it is only long-context coding assistance, one can wait for the API to open and integrate by call volume. Verifiable metric: compare output consistency and first-token latency between the two forms on tasks from one's own repository. Second, on how to use long context: the report says the model natively supports one million tokens and adopts a hybrid structure of SWA and sparse attention, so one should not by default stuff all history into a single request. First measure task completion rate and per-request cost at 128K, 256K, and 512K inputs, combined with retrieval or summary compression; if completion rate does not drop noticeably after compression, adopt a segmented-plus-retrieval approach and reserve long context for the few tasks that truly require associating across experiment records. Verifiable metric: the pass-rate difference and peak VRAM between compressed and full input on the same task set.
Limits and open questions
First, it must be made clear: as of this catalog check, the version Naive-N0.5-Flash does not appear in KimiSeek's model catalog, and the API availability mentioned in the report refers to Naive AI's own openness plan, not to anything already listed on this site. Second, the evaluation scores, training configuration, and NaiveRT and AutoWM case results in the report all come from the vendor's technical report and media relay, without third-party independent reproduction; some comparison targets have relatively new names, so it cannot be judged from this whether benchmark versions and test conditions are consistent, and cross-report horizontal comparisons require particular caution. Third, the MIT license is relayed by the source; the actual scope of authorization should be based on the license files shipped with the repository, and the fact that weights can be downloaded does not mean all components can be freely used commercially. Fourth, the input, output, and cache prices disclosed in the report are vendor quotes at the time of release, prices may change, readers should rely on the official pricing page, and should not use old quotes for budgeting. Fifth, the report does not state the actual effective length of the million-token context, the compatibility of the hybrid attention with third-party inference frameworks, or whether supporting offerings such as sandboxes and compute platforms will be open to the public.