Source: MarkTechPost — 2026-07-31
Summary
DeepSeek pushed deepseek-v4-flash-0731 into public API beta on July 31, 2026 — the same 284B-parameter, 13B-active MoE architecture as its April preview, at the same $0.14/$0.28-per-million-token price, but re-post-trained for agentic and coding work. The result: DeepSWE jumped from a preview score to 54.4, a 7.5x multiple, and Terminal-Bench 2.1 rose from 72.1 to 82.7 — within 2.3 points of Claude Opus 4.8's 85.0, at a fraction of the cost. Nothing about the model's shape changed; only the training data and process did.
Key Takeaways
- Architecture is frozen: 284B total parameters, 13B active per token, 1,048,576-token context, hybrid attention, DSpark speculative decoding — identical to the preview build.
- DeepSWE score: 54.4, up 7.5x from the preview. Terminal-Bench 2.1: 82.7, up from 72.1 (V4-Pro-Preview) — trailing Opus 4.8's 85.0 by just 2.3 points.
- Across nine published agent benchmarks, V4-Flash-0731 doesn't beat Opus 4.8 on any single one, but averages only 5.7 points behind it, at roughly 1/54th the blended token cost.
- The model now natively supports the Responses API format and is adapted for Codex-style tool-calling workflows — a serving-layer change, not a weights change.
- This is a clean case study in post-training ROI: no new parameters, no new context window, no new price — the entire gain came from re-training on better agent/coding trajectories.
Reel Script
Hook (~18s, 40 words): DeepSeek just proved you don't need a bigger model to get dramatically better at coding — they took their existing model, changed nothing about its architecture, and its agent coding score jumped seven and a half times over.
Core Concept (~70s, 160 words): Here's the mechanism that matters: DeepSeek-V4-Flash-0731 has the exact same "brain structure" as the preview version released months earlier — 284 billion total parameters, but only 13 billion of them actually activate for any given token, because it's a Mixture-of-Experts model. Think of MoE like a hospital with 284 billion dollars of specialist equipment, but each patient — each token — only gets routed to the 13 billion dollars' worth of specialists relevant to their case. That routing logic didn't change in this release. What changed is post-training: DeepSeek re-ran the fine-tuning process on better examples of agentic behavior and coding tasks, essentially giving the same brain better practice problems. No new parameters, no bigger context window, same price per token — and yet DeepSWE, a benchmark for real coding-agent tasks, jumped from a low preview score to 54.4.
Hands-On (~55s, 130 words): The concrete before-and-after: on Terminal-Bench 2.1 — a benchmark that scores an AI agent actually operating a command-line environment — the preview scored 72.1. The re-trained 0731 build scores 82.7. That's a 10.6-point jump from training alone. For comparison, Claude Opus 4.8 scores 85.0 on the same benchmark — so DeepSeek closed the gap to within 2.3 points of a frontier model, while pricing V4-Flash at $0.14 per million input tokens versus Opus's dramatically higher rate. Picture two bar charts side by side, same model architecture bar, but the "after re-training" bar nearly touching the frontier-model bar.
Takeaway (~25s, 55 words): If you're choosing a coding-agent backend on cost, this is the model to benchmark against your own workload before defaulting to a frontier model — you're likely leaving very little performance on the table for a huge price cut. Pull up the API docs and run your own eval before committing budget either way.