DeepSeek-V4-Flash-OR — an open model that formulates the math behind decisions
DeepSeek-V4-Flash-OR
Every hospital, airline, and delivery company runs on the same math. Given limited staff, trucks, or budget, what's the best way to use them? An open AI you can download was just trained to turn those plain-English problems into the exact math that solves them. It's called DeepSeek-V4-Flash-OR, from the Shenzhen Loop Area Institute.
An open model turns plain-English plans into solvable math
That math has a name: operations research, the math of making the best decision under limits, like routing forty trucks over two hundred stops for the shortest total drive. Normally a specialist turns that plain-English problem into a formal setup: the choices you can make, the goal, and the rules you can't break. Then a solver, standard math software that computes the provably best answer, crunches it. This model does the setup itself. It isn't new: the lab took an existing open model, DeepSeek-V4-Flash, and specialized it.
Its biggest edge is setting the problem up right
Its biggest edge is setting the problem up right, not guessing the answer
You'd expect an AI good at a math task to just get more right answers. But its single biggest edge is somewhere else: setting the problem up right, picking the choices, the goal, and the rules the way an expert would, instead of just landing on the same final number. On the makers' own test, first try, no examples, it scores fifty-nine on that setup, against the next-best model's forty-seven, and its own starting point's thirty-four. That's the sign it learned to build the problem, not memorize answers.
Immersion first, then a solver grades the drilling
Two training stages, and a real solver grades the second
So how do you teach a model operations research? Two stages, and the second one is the trick. First, immersion. The lab kept training the open base model on operations-research material, worked problems and solver-checked examples, until it absorbed how these problems are shaped: the choices, the goal, the rules, not just their answers. Second, drilling. The lab fine-tuned it on ten thousand problems where a real solver is the grader. The model writes the math, the solver runs it and checks two things: did the program return the known-best answer, and did the model set the problem up the way an expert would. So a solver, not a person and not another AI, decides right or wrong. The model is graded against ground truth. And immersion does the heavy lifting: with the drilling stage alone, the setup score is forty-nine. Add immersion first, same drilling, and it jumps to fifty-nine. Immersion is what taught it to build the right model.
Highest average of seven models, but not a clean sweep
Across four operations-research tests, first try, this model posts the highest average of the seven, seventy-two, ahead of GPT-5.4-Mini and Gemini-3-Flash, and eleven points over its own starting point. But this is not a clean sweep. It wins the average, the plain-English test, and the setup test, but loses two of the four, one to Gemini, one to GPT. Every model here was re-run by the same lab under one testing setup. A controlled comparison, but their comparison.
About 3x the training work on Huawei's chips — on the trillion-parameter sibling
About three times the training work on Huawei's chips
Here's how the whole project got built. The default chips for training at this size are Nvidia's. This lab built its training system for Huawei's AI chips instead. On the trillion-parameter big sibling of this model, a separate, much larger model called DeepSeek-V4-Pro, that system got about three times as much training work out of the same chips as the standard off-the-shelf setup. That system is the engine behind everything here.
An open model learned to formulate the math behind decisions
An open model learned to formulate the math behind everyday decisions
The scheduling, routing, and budgeting math that used to take a specialist, an open model you can download now does the hardest part itself: setting the problem up right. It got there by being graded against a solver, not a person. The weights are free to download, and the training recipe is public too, both under open, permissive licenses. You can pick it up today.
