Confirmed in a primary source 27 Sept 2026 Models

A Beijing company gave away a 309 billion parameter model and said models did most of the work of building it.

On September 27 NaiveAI, a Beijing startup founded by the Tsinghua professor Jifeng Dai, published the weights and inference code of Naive-N0.5-Flash under an MIT licence, which is the licence that attaches no conditions at all to commercial use. The company's own post says its models write the code, run the experiments and decide what to try next, that human researchers set direction and make the final calls, and that this is the path it is pursuing toward recursive self-improvement, which it names in those words.

The caveat that matters is that every number on the page is the company's own, produced with its own harness, and the source code that would let anyone check the headline speed figure is promised for October 12 and is not out yet.

Until now

Labs said models were starting to help with research, and the evidence was anecdotes and demonstrations.

Now

A company has published a model, a licence with no conditions on it, and a document saying models did the engineering and that the aim is to have them do the next generation.

The release itself would be a routine Sunday. Another large Chinese model, another set of benchmark charts, another Hugging Face page. What is not routine is the document beside it. Most labs describe models helping with research in the careful language of pilots and internal tools. This company describes an arrangement: models write the code, run the experiments, analyse the results and decide what to try next, and people set the direction and make the calls that matter. Then it gives two cases with counts. Six days and 151 trials to build an inference runtime, 63 of them kept and 71 thrown away. Four hundred hours and 15 rounds on a research problem the company says was outside its expertise, ending above the best published score on a public benchmark. Counts of failed trials are an odd thing to invent, because they make the process sound less magical rather than more.

The reason this is a card and not a model release is the last paragraph of the post, where the company names recursive self-improvement as what it is trying to do. Labs have argued about that phrase for a decade, mostly as a risk. Here it is a product roadmap, written by a company with a reported valuation of 1.4 billion dollars, attached to weights anyone on earth can download and keep. Note what cannot be taken back: the licence has no conditions, the file is mirrored the moment it is posted, and no later decision reverses that. Note also what is not yet checkable. The speed numbers are the company's own until the code appears on October 12, the benchmark comparisons are self-run, and a model's score on a benchmark it was trained to be good at proves less than its authors would like. A company has published a model, a licence with no conditions on it, and a document saying models did the engineering and that the aim is to have them do the next generation.

How it got here

Sep 22
Xiaomi releases the MiMo-V2.6 series, the top of the open-weight field on one composite index. Its earlier MiMo-V2.5 base model is what today's release is built on.
Sep 25
Anthropic publishes a nine-loop physics calculation a model finished unattended, for about 100 dollars of the budget. Yesterday's science card.
Sep 27
14:14 UTC, the Naive-N0.5-Flash repository is created on Hugging Face. The company's blog post of the same date names recursive self-improvement as its aim and gives trial counts for two worked examples.
Oct 12
The date the company itself puts on publishing the runtime source code and benchmark scripts that would let anyone reproduce its speed figures.
Who has to react
Every lab that has argued the frontier cannot be reached without its own scale, since this one started from somebody else's base model; Xiaomi, credited by name and now upstream of a competitor's release; the American firms whose coding tools were used as the test harness; Chinese regulators, who have a 309 billion parameter model under a licence with no restrictions leaving the country by download; the export control officials whose instrument is hardware rather than files; StepFun, which promised weights on October 15 and is now being measured against a company that just published without promising; enterprise buyers already routing a majority of their tokens to open weights; and anybody who has spent a decade arguing that recursive self-improvement is a risk rather than a roadmap.

The evidence

ObservedSunday September 27, 2026. The Hugging Face repository NaiveAI/Naive-N0.5-Flash was created at 14:14:32 UTC and last changed at 19:13:58 UTC the same day, which is where the date comes from.

The whole observation, as recorded

The model is 309 billion parameters with 15.5 billion active per token, 48 layers, a native one million token context window, and no full-attention layers anywhere in it: 39 sliding-window layers and 9 sparse layers. It is built on Xiaomi's MiMo-V2.5 base model, which the company credits by name, and it was trained on 3.25 trillion further tokens. Weights and inference code are under the MIT licence. The API price is 10 cents per million input tokens and 40 cents per million output. The company's blog post of the same date is titled "Naive-N0.5-Flash: Building Frontier AI with AI" and states: "At NaiveAI, we broke with human-centered R&D from day one. We put AI models to work on AI R&D itself. They write code, run experiments, monitor progress, analyze results, and iterate. Human researchers set direction, define constraints and criteria, and make critical decisions." It says the company runs "close to ten million sandboxes each week, with 100,000 active concurrently at peak." Two worked examples carry numbers. NaiveRT, the inference runtime, was "built in six days by human researchers working with AI models, across 151 documented optimization trials": 63 adopted, 71 failed validation or were rolled back, 17 explored alternatives. It reaches a peak single-stream rate of 2,122 tokens per second on eight processors and cuts a full speculative decoding round from 12.3 milliseconds under an existing open-source system to 3.4, a 72.4 percent reduction. The second example, AutoWM, was a research task in a field the company says it had no expertise in: a researcher set the objective, the compute budget and the evaluation protocol, and after "400 hours of cumulative research and 15 major experimental rounds" the resulting model scored 77.43 on a public video benchmark against a previous best published score of 73.64. The post's own conclusion: "This is the path we are pursuing toward recursive self-improvement (RSI): AI models taking on the research that produces the next generation of models, with each generation able to take on more of that research." Two things are not established. The benchmark tables are the company's own, run on its own harness, using Claude Code version 2.1.207 as the test rig. And the NaiveRT source code and benchmark scripts that would let anyone reproduce the speed figures carry the line "The above content will be available by Oct, 12th." The Information has reported a 1.4 billion dollar valuation for the company.

Established

NaiveAI published the weights and inference code of a 309-billion-parameter mixture-of-experts model with 15.5 billion active parameters and a one million token context window under the MIT licence on September 27, 2026.
The model is built on Xiaomi's openly published MiMo-V2.5 base model and was trained on a further 3.25 trillion tokens.
The company states in its own published post that AI models write code, run experiments and iterate on its research, that human researchers set direction and make critical decisions, and that it is pursuing recursive self-improvement.
The company gives 151 documented optimisation trials over six days for its inference runtime, 63 adopted and 71 rolled back, and 400 hours across 15 experimental rounds for a world model that scored 77.43 against a previous published best of 73.64.
The company runs its own compute environment for this, which it says serves close to ten million sandboxes a week with 100,000 active at peak.

Not established

That the benchmark results hold up. Every figure is the company's own, measured on its own harness, and no independent evaluation exists.
That the speed claims are reproducible. The source code and scripts that would let anyone check them are promised for October 12 and are not published.
That models designed the model. The post says researchers defined objectives and picked among the candidates the models implemented, which is assistance at scale rather than autonomy.
That anything here is recursive yet. One generation built with help from an earlier generation is a step, not a loop, and the company has published one generation.
That the valuation figure is confirmed. It comes from reporting, not from a filing.

Why it weighs this much

Signal strength

CostHigh

CommitmentHigh

ReversibilityHard

PersistenceStructural

Cost is high. Training 3.25 trillion tokens at a million-token context is real money, and giving the result away under a licence with no commercial conditions forgoes the revenue that would normally pay it back. Commitment is high because the claim is published under the company's name in its own technical write-up, with trial counts and failure counts that invite checking, and because a dated promise to publish the source code on October 12 is a hostage. Reversibility is hard, which is the point of an open-weight release: the file is copied within minutes and no later decision pulls it back. Persistence is structural because an open frontier-class model aimed specifically at coding and machine learning research does not stop mattering when the news cycle moves, and because the stated aim, models doing the research that builds the next models, either compounds or does not, and we will be able to see which.

What to watch

strengthens The NaiveRT source code and benchmark scripts are published by October 12, 2026, and an independent party reports reproducing either the 2,122 tokens per second figure or the 77.43 benchmark score, by December 31, 2026.
weakens October 12 passes with no source code published, or an independent evaluation places Naive-N0.5-Flash outside the field of open-weight coding models it publishes itself against, by December 31, 2026.

The words

Open weights: The trained numbers inside the model are published, so anyone can download and run it on their own machines without asking permission.
MIT licence: One of the shortest and most permissive licences in software. Use it, sell it, change it, no conditions beyond keeping the copyright notice.
Mixture of experts: A design where only a small slice of the model runs on any one word. Here 15.5 billion parameters out of 309 billion do the work each time, so it costs far less to run than its size suggests.
Recursive self-improvement: The idea that a system good enough at building systems can build a better version of itself, and that version a better one again. The company uses the phrase about its own plan.

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.