Small AI models now beat flagships because most of the gains in 2026 come from post-training, not size. DeepSeek improved its V4-Flash model from 40 to 50 points on the Artificial Analysis Intelligence Index without changing its architecture or parameter count, which put it ahead of its own larger V4-Pro at 44. For teams paying per token, that changes how you should pick models.
This explainer covers what happened, the three techniques behind it, and a practical way to choose a model for your own work.
What happened in 2026
Three releases made the point clearly.
DeepSeek-V4-Flash-0731 (July). According to The Batch, DeepSeek says the jump came from a new round of fine-tuning on an unchanged architecture. The model has 284 billion total parameters but only 13 billion active per token. Terminal-Bench 2.1 rose from 61.8% to 82.7%, and its cost per Intelligence Index task was about $0.03.
Inkling Small (same week). Thinking Machines' smaller model scored 40 on the same index, within one point of its flagship Inkling, with less than a third of the parameters, per The Batch.
DeepSeek-V4.1-Flash (September). DeepSeek's release note describes a 552 billion parameter model with 8 billion active parameters for input and 16 billion for output, which it says beats V4-Pro. DeepSeek then routed all V4-Pro requests to V4.1-Flash from 14 September, at Flash prices.
| Model | Total / active parameters | What changed | Result |
|---|---|---|---|
| DeepSeek-V4-Flash-0731 | 284B / 13B | Post-training only | Index 50 vs V4-Pro's 44 |
| Inkling Small | Under a third of flagship | Smaller model | Index 40, one point below flagship |
| DeepSeek-V4.1-Flash | 552B / 8B in, 16B out | New architecture, bigger RL | Replaced V4-Pro |
A caution on the numbers: benchmark scores come from DeepSeek and from Artificial Analysis, and real workloads can differ. Test on your own tasks before switching.
The three techniques behind it
1. Post-training does more work than people expect
Pretraining gives a model broad knowledge. Post-training teaches it how to use that knowledge: following instructions, calling tools, reasoning step by step, and knowing when to stop. DeepSeek's V4-Flash update is a clean example because the base model stayed the same and only post-training changed.
The Batch describes DeepSeek's recipe in two stages. First, separate specialist models for math, coding and agentic tasks were trained with supervised fine-tuning and then reinforcement learning using Group Relative Policy Optimization (GRPO). GRPO scores several answers to the same prompt against each other, which avoids training a separate value model. Second, the specialists were merged into one model.
2. Distillation moves skills from many models into one
That merge step used on-policy distillation from more than ten specialists, per The Batch. In plain terms, the final model generates its own answers and learns to match what the relevant expert would have produced. It ends up with the strengths of each specialist in a single set of weights you can serve cheaply.
3. Mixture-of-experts keeps the bill small
A mixture-of-experts (MoE) model contains many sub-networks but only activates a few for each token. V4-Flash holds 284 billion parameters of knowledge yet does roughly the compute of a 13 billion parameter model per token. That is why it can be both capable and cheap. The Batch also reports that, at full input length, it needs about 10% of the compute and 7% of the key-value memory of DeepSeek-V3.2.
Why this matters for cost
Agents burn through tokens. An agent that triages bugs or reconciles invoices may run thousands of steps a day, so cost per task decides what is worth automating. The Batch makes the point directly: near-proprietary quality at well under half the cost per task makes always-on agent work practical. V4-Flash also ships under the MIT license, and a 3-bit quantized version runs on a machine with 110 GB of memory, so you can keep data on your own hardware.
There is a catch. Smaller models are not always cheaper per task. Some reason at great length to reach their scores, which eats tokens. Measure cost per completed task, not price per million tokens.
How to choose a model for your workload
- Write 30 to 50 real test cases from your actual work, with clear pass or fail checks.
- Run two or three candidates, including at least one small or open-weight model.
- Track pass rate, cost per completed task and latency. Price per token alone misleads.
- Route by difficulty. Send routine steps to the small model and escalate hard cases to a frontier model. Our enterprise agent routing playbook shows how.
- Re-test every quarter. Rankings now change in weeks, as DeepSeek's July and September releases showed.
For a wider view of the trade-offs, see our guides on open-weight vs API models and price-performance for coding models.
Limits worth knowing
Index scores compress many tasks into one number, and vendors pick the benchmarks they publish. DeepSeek has not disclosed its training data or knowledge cutoff for V4-Flash, per The Batch. Model retirements also bring risk: when DeepSeek routed V4-Pro traffic to V4.1-Flash with a few days' notice, any prompts tuned to the old model's behavior effectively changed overnight. Pin model versions where your provider allows it.
Our take: for most business workloads in 2026, the right default is the smallest model that passes your own tests, with a frontier model as the fallback. Bigger is no longer a safe proxy for better.
Sources
- The Batch: DeepSeek pushes the frontier again
- DeepSeek: DeepSeek-V4.1-Flash release note
- The New Stack: DeepSeek's smaller model outperformed its own flagship
Researched and written by the Vocxa Blogs editorial team with AI-assisted drafting. Facts checked against the linked sources as of 21 November 2026.
