Jev was asked to take the throne: StartLux.

On September 30, Shanghai AI, StartLux, released a fully open source decision model. StartLux-DecisionA one-time launch of 0.8B, 2B, 4B, 9B, 27B, 5. According to the evaluation published by the team, its flagship 27B version received 63.88 points on Decision Index 0.2.1, more than Jeff 1.13 (57.91), the first on the list.The new king of the decision-making model led to the throne.I don't know.

More interesting is the rate of R & D:The entire model was only 3 days to complete research and validation.I don't know. Behind this is the Auto Research line built by StartLux -- officially, it's one of their core paths to advance RSI.

Let's start by saying, this is a team test, not an official list.

This one must be clear, otherwise it is easy to be led over the title.

StartLux's grades are:Using the official evaluation tool of the Decisión Index, the team self-measured against the public listshot completed on September 28, 2026I don't know. In other words:

• They're using the checklist assessment tool and the same day's comparison data.
• But...Their grades are not on the official list.It's also written in the official article that "not that it's officially certified."
• Yeah.Top score on the public list.(2026-09-28)

So the exact answer is:"StartLux-Decisión has surpassed the top of the current public list by team self-measured."I don't know. It's not demeaning — it's self-measured that means strength, but the reader needs a scale. All the numbers below are taken with the default premise of "team self-measurement".

What is it: a decision model that does not generate text?

Start by understanding the fundamental difference between the StarLux-Decisión and the general large model. It's not a generation model.Do not write natural languagesI don't know.

That's how it works: give it one.StatusAdd a groupQuestions— the selection of the theme (one from several options), is not, or is graded. The state can be text, JSON or picture, up to 256 K token.Returns the probability of each question to each optionI don't know.

The point is in the last sentence:It doesn't produce anything.I don't know. The answer is read directly from the option letter after a forward transmission. So a short question takes a few milliseconds and the probability can be used directly as confidence.

This solves a particular kind of work in a smart body:HF, clear candidate setThe judgment. For example, the passenger service order came — “the order is repeatedly withheld but still shows no payment” — and the same information can ask three questions at the same time: which team was given it, whether it is being processed today, and how much impact it has had. The traditional approach is to allow the generation model to write a piece of it, which is slow and easy to interpret errors; one request from the StateLux-Decisión can complete the three judgements together.

Report cards: 63.88 vs. 57.91,38 out of 31 won

A comparison of Decision Index 0.2.1 (both 2026-09-28 snapshots):

• StartLux-Decision-27B:63.88
• Jev 1.13:57.91 (first place)
• Rune 26B-A4B: 57.44
• Decider Chat 31B:57.33
• AutoJev-27B: 56.40
• StartLux-Decision-9B: 58.639B is ahead of Jeff.

27B higher than Jev 5.97 minutes, of the 38 benchmarks contained in the index,There are 31 higher scores.I don't know.

But please note the official quote:"The combined score does not mean that all capabilities are stronger."I don't know. This sentence is worth reading.

Where to win, where to lose.

In the area score for Decisión Index, 27B leads in four areas, but loses to Jev:

• Language understanding74.5 vs Jev 62.0 (high 12.5)
• Search and classification:66.8 vs 55.4 (high 11.4)
• Tools and automation:82.2 vs 75.1 (high 7.1)
• Art and human preferences:47.9 vs 37.7
• Knowledge and reasoning:44.3 vs Jev 51.4 (Later 7.1)

This distribution is very illustrative. StartLux-Decision Strong“understanding conditions, distinguishing among candidates, choosing the next course of action”— It happens to be the three types of capabilities that are most needed in the day-to-day implementation of intelligence; it is not superior to purely intellectual reasoning. So the conclusion is not "who's crushing who," but...The selection depends on what kind of judgment your intelligence is most likely to encounter.I don't know.

A similar division exists on individual items. For example, API-Bank, which does not lead, Home applications pull from 52.3 in Jev to 77.3, and POP909 from 15.9 to 50.6. These are the scores of Decisión Index that have been corrected by random baseline.Could not directly be considered as the accuracy or completeness of the original version of the baselineI don't know.

Seven public tests: 91.82% average, few mistakes 8

In addition to Decisión Index, the team ran away from Shanghai's artificial intelligence lab to launch the Inter-Design.Seven sets of tests(JevBench EASY/Original/Hard, Typed Decisions, ToolaCE, AG News, WildjailBreak):

• StartLux-Decision-27B average accuracy rate 91.82%
• Jev 1.13:88.74%
• Inter-Decisión-4B: 90.02%

Publicly, JevBench has 231 questions, three answers. 208 / 199 / 201 Question - 27B vs. 9 more than Jeff. In the hardest JevBench-Hard (111 open dilemmas), 27B error 23, compared to Jev 31Question 8I don't know.

Here, too, there is a lot to be noticed: the public JevBench’s 208/231 is just an open question.It's not an official score of classified tests, speed and cost.; additional training data used the open training set of 14 of these benchmarks and the corresponding test questions were filtered out (official label).

Speed: milliseconds, all problems counted forward at once

The value of the decision model is half right, half fast. In a short request test for card H200, BF16, local HTTP,Three questions at a time.time-consuming:

• 0.8B:12.2 ms
• 2B:15.5 ms
• 4B: 26.0 ms
• 9B:35.7 ms
• 27B:102.3 ms
• 35B-A3B (MoE): 52.5 msEvery token in 35B only activates about 3B, twice as fast as 27B.

When you answer only one question, 4B on average. 14.7 msI don't know. Optimization within the same model is more indicative of engineering gains: 4B CUDA Graph was 90.3 ms when it was not enabled, down to 26.0 ms when enabled- More than three times the promotion depends on engineering. The project also uses quick linear attention algorithms and consolidates multiple issues into a forward calculation to reduce double calls.

As a reference, Jev 1.13 on the three requests of the same kind, the service end reported by its API gateway is 64.0 msI don't know. However, the authorities have restrained themselves from reminding:Hardware and service depots vary from one party to another and cannot be used as a basis to conclude that hard hardware is the same speedand these milliseconds are not the time to complete the entire web-page task (both input length, page loading, tool execution). Rather, this self-contained expression is the more credible part of the publication.

Field presentation: from browser operation to game

The official presentation covered several directions, with the core being "models for decision-making, implementation systems for action":

Browser/ Web Operation: Each step asks two questions — which controls are the next operation and whether the task is completed. In the demonstration, the filtering list was completed for "Buy the cheapest, free of charge, 8 AA batteries sent home" and the designated members were invited to the Design team and set up as Editor. Attention to official statements:Controlled environmentThe following is a continuous operation.It's not the same as free entry on any website., the input box also completes the selection through the candidate proposal.

Chess: StartLux-Decisión-27B and Jev 1.13 compete, neither search, and the code never overwhelms the selection of models (Stockfish only draws assessment bars). It's just a matter of time.256, 27B, 58.4%.I don't know.

Game: Mario (simulate 11 candidate actions first, write the results into text before selection), StarCraft II ( ask three questions every 12 seconds of game time: what to build, what to give priority to, what to attack/defend/collect), DOOM (each 4 Tic judge targets, fire, move), JevBall (one request for separate action for several players near the ball). In the same vein:These are not a uniform Direct Read Screen test., there are scripts for perception and execution, and the model is only responsible for the part of decision-making defined by the interface.

And there's one more detail that's interesting -- "Hanna, have you seen Jeff today?" There's Hannah and Jeff in that sentence. The system has to judge.He's been questioned about Hannah.Instead of having another actor respond because of Jeff. It's taken separately to determine whether the model understands the conversation or only matches the name.

Five sets of specifications + MoE + Quantified, locally deployed friendly

StartLux-Decistion provides a five-tier dense model (0.8B / 2B / 4B / 9B / 27B), plus one 35B-A3B Mixed Expert (MoE)Version. October 3rd.Picture input and 256K token contextEach one of them can be read as evidence.

Local deployment is the focus: llama.cpp is available for each tranche GUF Files, including BF16 to retain original accuracy, and two quantitative formats Q8 0, Q4 K M,15 GGF entriesI don't know. Take the example of 4B:

• Q8 0 Files 4.48 GBIn 231, on JevBench with the original right HeavyDecision-making coherence rate 100%(Same answer for 204)
• Q4 K M presses 2.71 GB98.3%, yes 201

Perhaps you should pay particular attention.: Code uses Apache-2.0, butModel weighting with CC BY-NC 4.0 - Research and non-commercial only, commercial use subject to a separate StartLux Labs authorizationI don't know. This is different from the “full open source” propaganda, which must be confirmed before commercialization.

The interface is compatible, Jev. Probability can be diverted.

There's a smart design at the access level: StartLux-Decisión with TypeSafe /v1/systemone Interface format,Client for Jev can be used as it is, business processes retain the current `state plus problem' mode of call and change the service address. The interface (llama.cpp load model with accompanying decision-making services responsible for organizational issues and reading out options) is also maintained for GGF deployment.

Another mechanism of concern isProbability diversionSince each candidate has a probability, the system can directly implement decision-making in a “validated routine scenario”; when the candidate is difficult to distinguish, is in inadequate condition or has a higher operational risk, it can be followed up with additional observations, submitted to a large model analysis, or requests for manual confirmation.

The government is also giving a real warning:Probability output does not equal the real success rate, and the diversion threshold is aligned with operational data calibration.; payment, deletion, authority to modify such operations shall remain subject to the original authorization and confirmation rules - - The decision model is the next step for the system to judge, not to replace these constraints.

Auto Research and RSI behind

Three days to make a five-stage model, depending on StartLux. Auto Research Development Pipeline(Constructed by co-founder and CTO Guo Guo Guo Jiu) According to a previous interview with the machine heart, in this approach, AI did not only carry out pre-set experiments, but also...Analysis of failures based on the previous round, presentation of new assumptions, participation in deciding what to look at next; while research objectives, evaluation criteria and key rules remain open.

In the system division of StartLux,StartLux-27B to carry the generic capability layer, Decisión model focused on the high frequency decision layerAnd above it is Agent and Memoory, the bottom is supported by quantification, reasoning systems and hardware. So Decisión is not an isolated release made in pursuit of hot spots, but a piece of this architecture that was built up as planned.

Guo Quan defined Auto Research as the core element for achieving RSI: First, to involve AI in research and model improvements, then to further request “improved AI, even the ability to improve AI itself has been enhanced”. StartLux-Decision is a concrete result of this line.

What about this?

Put Jev on the throne.A Chinese team at the bottom of Agent's Decision Level.I don't know. The idea that Jev brings — using a dedicated model to take on high frequency, to stand for a clear judgement rather than allowing the generation model to be both a referee and an athlete — is proving to be a real need. StartLux uses an open source, a fifth set of specifications, local deployment and a Jev compatible interface to make the road more “landable”.

But three reminders still stand:

• It's a team test.Not certified on the official list.
• All ahead, all ahead.It still lags behind in intellectual reasoning, Jeff. It depends on the distribution of tasks.
• The weight is CC BY-NC 4.0., "Full open source" only for non-commercial scenarios, with commercials authorized separately

In any case, it is worth remembering that a relatively young Shanghai team, three days of making a five-stage decision model and surpassing the current top of the public calibre.Agent's “judgement level” is becoming an independent and deserving technology track.

The decision model is not intended to replace the big model, but rather to make less of the trivial judgements it should not make.