DeepSeek V4 Flash Gained 47 Points Without Changing Architecture—Here's the Real Lever

Yesterday (July 31), three things happened in the AI world one after another.

In the morning, Zhipu GLM Coding Plan reopened subscriptions.

Previously, they released quotas at fixed times every day, and I never managed to grab one. Now there's no need to scramble.

I thought it was good news, but when I saw the official subscription prices, I was dumbfounded.

Lite 118, Pro 538, Max 1078.

The highest tier is more than 3 times more expensive than the old version, and the usage quota seems to be even lower.

Immediately after, in the morning, I saw news about GPT 5.6 price cuts.

OpenAI is doing the opposite: the Luna series of GPT-5.6 has seen price cuts of up to 80%.

Price cuts abroad, price hikes at home—the plot is a bit surreal~

Then I came across the release of DeepSeek V4 Flash 0731 official version.

Looking at the benchmark scores, they've surged across the board, surpassing GLM-5.2 in several metrics, with no price increase at all—still maxing out the price-performance ratio!

Amidst the excitement, I stared at the official benchmark comparison chart that evening for a long time.

DeepSWE, 7.3 → 54.4. Terminal Bench 2.1, 61.8 → 82.7. Cybergym, 38.7 → 76.7.

You have to know that 0731 and Preview share the same architecture, same scale, only the post-training was redone.

Without changing the architecture or adding parameters, how did it get so much stronger?

I only knew that reinforcement learning and post-training could improve a model's ability to do things, but can post-training alone really improve it that much?

I researched this question and found that the answer is far more complex—and far more interesting—than just 'post-training is awesome'.

What 0731 Did: The Official Word is Silent, But There Are Clues

What exactly 0731 changed—the data recipe, the training ratios—has not been disclosed by the official sources.

But it's not without clues.

The V4 technical report describes the complete post-training methodology for this generation, and 0731 most likely comes from this system.

Understanding it is the prerequisite for everything that follows.

V4 Flash is an MoE (Mixture of Experts) model: total parameters 284B, only 13B activated per token—a bunch of 'experts' inside, only waking up a relevant small group to work each time, hence it's both powerful and cheap.

Its post-training is divided into several steps:

Step 1: Train experts separately. Code, math, and tool use are each trained independently: first SFT (Supervised Fine-Tuning, providing standard answers), then GRPO (Group Relative Policy Optimization, scoring based on outcome quality).

Step 2: Distill and merge. Using On-Policy Distillation (where the student model aligns with the teacher token-by-token in the states it actually visits), merge 10+ experts back into a single model.

Step 3: Be your own judge. For tasks where right or wrong cannot be judged, use a GRM (Generative Reward Model) to let the model score itself.

Additionally, a DSec sandbox platform was built: hundreds of thousands of isolated environments in a single cluster, where the model runs code and invokes tools for real.

The key insight is: Post-training is no longer just 'feeding some conversation data to adjust speaking style'; it's an industrial system with expert division of labor, distillation, a judge, and a training ground.

As for which knobs 0731 tweaked within this system, that's still unknown.

Is Post-training 'Teaching New Things' or 'Cashing in Old Abilities'?

This is the most critical question in the entire investigation.

First, let me lay out my previous naive understanding: pre-training pours in knowledge, post-training teaches how to do things—a bookworm is trained into a skilled worker.

The general direction isn't wrong, but it's too vague. Currently, there are two schools of research:

School 1: Mainly 'Cashing In'. A NeurIPS 2025 paper used pass@k for experiments: let the model attempt the same problem k times, passing if it gets it right once.

The result was that problems solvable by the model after RL were basically solvable by the base model, just requiring more attempts.

That is to say, RL didn't teach new solution methods; it turned 'occasionally getting it right' into 'getting it right the first time'. The ability was always in the base model; post-training improved the probability of the correct strategy being selected.

School 2: RL Can Also Expand Boundaries. NVIDIA's ProRL found that with sufficiently long training, the model could solve problems that the base model couldn't solve no matter how many attempts it made; and the expansion was more pronounced in domains where the base model was weaker—those areas hadn't been 'squeezed dry' by pre-training yet.

My understanding is: The default is cashing in, but when the base model hasn't been squeezed dry and training is long enough, it can also expand boundaries.

Distillation is an exception; both schools acknowledge it can transfer new behaviors—the teacher might inherently be another, stronger model. Having a small model directly learn the long chain-of-thought and problem-solving abilities of a large model is already a mature practice in the industry.

So Why Did the Scores Jump So Dramatically?

Even if we accept 'mainly cashing in'—DeepSWE increased 7-fold, where does such huge leverage come from?

Actually, this might be a cumulative effect.

Suppose an Agent task requires 100 consecutive operations (reading files, modifying code, running tests, reading error messages... hundreds of tool calls are normal), and the single-step accuracy is 90%, which sounds quite high.

But the probability of all 100 steps being correct is 0.9 to the power of 100:

• Single-step 90%: Overall perfect probability 0.0027%, basically impossible to succeed
• Single-step 99%: 36.6%
• Single-step 99.9%: 90.5%

A '47-point jump' on the books is, at the bottom, a tiny improvement in single-step stability, exponentially amplified by the long chain.

This is the most ruthless leverage of post-training: it doesn't teach new things, it just cashes in existing abilities into stable strategies—and on long-chain tasks, the value of 'stability' is exponential.

Code and Agent tasks are particularly suited for this kind of training:

Rewards are verifiable. Code can run tests; right is right, wrong is wrong.

Correction at the error site. On-Policy training covers the exact mistakes the model itself actually makes.

Credit assignment has solutions. Cursor's Composer 2.5 inserts a local prompt at the step where an error occurs, uses the prompted model as a teacher, and writes the correct behavior back via distillation, only fixing the erroneous segment.

Post-training itself is also still scaling: one study used 400,000 GPU hours to fit a sigmoid growth curve; Cursor's Composer 1.5 scaled up RL by 20 times on the same base model, post-training compute exceeded pre-training, and capability was still growing.

A Cold Shower: Not All the Score Gains Are the Model's Credit

The story is very smooth up to here, but it would be dishonest to stop now.

Mixed into the benchmark score increases are many things unrelated to the weights.

The most devastating experiment: Using completely random rewards for RL still improved Qwen2.5-Math by 21.4%—random rewards can't teach anything; it's the GRPO mechanism naturally amplifying the base model's priors.

Looking solely at 'score increase after RL', you can't distinguish whether the gain is capability or algorithmic bias.

Now look at the harness (the Agent's scaffolding, which determines what the model sees, what tools it uses, and how it retries on errors). Kimi K2.5 ran Terminal Bench under three harnesses: 43.2, 47.3, 50.8.

The same model, just changing a layer of scaffolding, differed by 7.6 points.

Viewing the battle reports with this perspective changes the flavor.

Composer 2's SWE-bench Multilingual, according to Moonshot's official statement, went from 73.0 to 73.7, almost a tie; the biggest gain was on CursorBench, Cursor's own benchmark, which cannot be externally reproduced.

DeepSeek's official Agent scores use a 'not yet publicly available DeepSeek Harness' 👀

Third-party Artificial Analysis independently tested and confirmed real progress (Intelligence Index about 10 points higher than Preview), but also pointed out its verbose output, generating 210 million tokens across the entire evaluation—part of the score is piled up with inference tokens.

There's also reward hacking: Cursor itself disclosed that the model discovered high-risk modifications were easily penalized, so it learned to use 'clarifying questions' to avoid modifications and collect safe points for free. The model is pleasing the reward function, not necessarily getting stronger.

So the next time you see a benchmark battle report, ask three questions first: Which version of the benchmark? Whose harness ran it? What inference tier was used?—Each variable is worth several points.

My Current Understanding: Capability is a Five-Layer System

After finishing the research, the single coordinate of 'model capability strength' in my mind turned into a layered diagram:

Using this framework to look back at yesterday's three events:

The price war is just the surface; the real arms race is in post-training—after pre-training scaling slowed down, post-training is the cheapest path to improve capability.

With the same hand of cards, whoever has a higher cash-in rate and whose training ground is closer to the real world wins.

Cursor has tightly integrated the model, harness, training environment, and real user feedback into a closed loop, producing a new checkpoint as fast as every 5 hours—the 'God of Post-training' isn't backed by a magical algorithm, but by an entire system.

DeepSeek has a different play: passing the technological dividend on to developers.

Liang Wenfeng once said, 'Our principle is not to lose money, nor to make exorbitant profits, only to earn a reasonable profit.' 🫡

Now, most models worldwide have entered DeepSeek's killing zone; those with inferior capability and high prices can only wait to be slain 😀

Back to the opening question: Same scale, why suddenly so much stronger?

My answer: No magic.

The strong 284B base model already had a massive amount of uncashed capability compressed within it; post-training is about releasing them layer by layer.

This is also why I'm looking forward to V4 Pro—Flash's base is 284B, Pro's is 1.6T.

If the 'cashing in' logic holds, what's compressed in the Pro base is only more.

Liang Sheng, hurry up.