DeepSeek V4 Flash 0731 — When Post-Training Beats Parameter Count

DeepSeek moved V4 Flash out of preview on July 31 with a re-post-trained checkpoint that the company claims outperforms its own 1.6T-parameter V4-Pro Preview on every published agentic benchmark. The architecture is identical to the April preview — 284B total parameters, 13B active per token, 1M-token context window. The only thing that changed is the post-training pipeline. At $0.14/$0.28 per million input/output tokens with MIT-licensed weights, this is the cheapest model in the capable-agentic tier by a wide margin, and it dropped without requiring a single line of code change for existing users.

TL;DR

• What: DeepSeek-V4-Flash-0731 ships same 284B architecture, re-post-trained only, vendor claims it beats the 1.6T V4-Pro Preview on agentic benchmarks
• Price: $0.14/$0.28 per million input/output tokens — 4× to 35× cheaper than comparable agentic-capable models
• Signal: Post-training pipeline quality is outpacing parameter scale as the performance lever that matters
• Action: Benchmark it this week on your own tasks. At this price, the cost of evaluating is near zero

DeepSeek V4 Flash 0731 — What Happened

At 07

UTC on July 31, 2026, DeepSeek published the V4-Flash-0731 checkpoint on Hugging Face under an MIT license and simultaneously updated the deepseek-v4-flash API endpoint. If your harness was already calling that endpoint, you got the new model automatically — no migration, no new model string, no re-integration.

The technical specifications are unchanged from April: 284B total parameters in a Mixture of Experts architecture with 13B parameters active per forward pass, a 1M-token context window, and native support for the OpenAI Responses API format. DeepSeek’s model card is explicit that this is “only re-post-trained” — same pretraining data, same architecture, different post-training recipe.

What changed is the claimed performance. DeepSeek states that Flash-0731 now outscores V4-Pro Preview — a model with 1.6T total parameters and 49B active — on every agentic benchmark they publish. The company positions this as a Flash-first strategy: ship the efficient model to capture developer integrations, then follow with the Pro-tier flagship later. V4-Pro remains in preview as of August 5 with no GA date announced.

The Responses API native support and explicit Codex adaptation deserve attention. If your agent framework already speaks the Responses API, this is a configuration toggle. The endpoint also works as a drop-in for any OpenAI-compatible client, which means switching a cost-sensitive agent loop from GPT-5.6 Luna or Claude Sonnet 5 to Flash-0731 is a model string change.

For self-hosting, the Q8 quantized checkpoint weighs in around 162 GB on disk. That puts it within reach of a multi-GPU server or a high-RAM single node, though inference RAM requirements will vary by serving framework.

All published agentic benchmark scores for Flash-0731 are DeepSeek’s own. No independent evaluation from Artificial Analysis, METR, or SWE-bench Verified has been published as of August 5, 2026. Treat capability claims as hypotheses until you run your own evals.

Why This Matters

The pricing gap between Flash-0731 and every other model with documented agentic tuning is not incremental — it is structural. Claude Sonnet 5 runs $2/$10 per million tokens during its promotional window through August 31. Grok 4.5 sits at $2/$6. GPT-5.6 Luna, the cheapest competitor with serious agentic credentials, is at $0.20/$1.20. Flash-0731 at $0.14/$0.28 is between 4× and 35× cheaper on output tokens depending on which model you compare against.

Output tokens dominate agentic workload costs. An agent that reasons through multi-step coding tasks generates substantially more output than input over a session. The difference between $0.28 and $1.20 per million output tokens compounds fast when your agent is running 50 or 500 tasks a day. For teams already watching their agent bills climb — and if you’re running production agent loops, you are — this repricing changes the math on which tasks are economically viable to automate.

But the more interesting signal is not the price. It’s what the price implies about where capability gains come from now. DeepSeek did not build a bigger model. They did not train on more data. They took an existing architecture and improved the post-training pipeline — the reinforcement learning, the instruction tuning, the alignment work that happens after pretraining. And they claim the result beats a model with 5.6× more total parameters on the tasks that matter most for developer tooling.

This pattern is not unique to DeepSeek. Google’s Gemini 3.5 Flash outperforms Gemini 3.1 Pro on agentic benchmarks while running 4× faster and costing 25% less. The Flash tier across providers is converging on a playbook: smaller active parameter counts, aggressive post-training investment, and pricing that makes Pro-tier models hard to justify for latency-sensitive workloads. The pretraining budget arms race is not over, but the marginal return on adding parameters is clearly declining relative to the return on post-training quality.

If you’re using a model router like Factory Router, Flash-0731 is a natural addition to your candidate pool. At $0.28/MTok output, even if the router selects it for only a fraction of tasks, the blended cost reduction on agentic workloads could be significant.

For teams evaluating self-hosting economics, the MIT license and manageable quantized size change the calculation too. A 162 GB Q8 checkpoint on your own infrastructure means zero per-token cost after hardware amortization. Compare that to running Gemma 4’s 26B MoE (3.8B active, Apache 2.0) at the edge — Flash-0731 sits in a different weight class but serves the same philosophical argument: open weights at competitive capability levels make API lock-in optional, not inevitable.

The catch — and it is a real one — is verification. DeepSeek publishes benchmark numbers but the agentic evaluation ecosystem does not yet have standardized third-party reproduction for these claims. Terminal-Bench 2.1, DeepSWE, and Cybergym scores are vendor-stated. I’ve seen enough vendor benchmarks that quietly cherry-pick evaluation conditions to know that “beats Pro on every benchmark” deserves an asterisk until someone independent confirms it.

The Take

I would not route production-critical agent workloads to Flash-0731 today based solely on DeepSeek’s published benchmarks. But I would absolutely benchmark it this week, because the cost of being wrong at $0.28 per million output tokens is negligible and the cost of ignoring it if the claims hold up is real.

The deeper story here matters more than any single model release. Post-training quality is becoming the primary lever for capability gains in the Flash tier. DeepSeek is running the same playbook that made Gemini 3.5 Flash competitive with Pro-tier models: invest in the post-training pipeline, not the pretraining budget. If this pattern holds — and the evidence across multiple providers says it does — then the question for your agent cost model is not “which model has the most parameters?” but “which provider has the best post-training pipeline for agentic tasks?”

The alias retirement deadline for legacy DeepSeek endpoints makes the timing pointed. If you are already migrating off deepseek-chat and deepseek-reasoner before the July 24 cutoff, you are touching your model configuration anyway. Add Flash-0731 to your eval suite while you are in there. Run it against your actual workloads, not synthetic benchmarks. At this price point, the evaluation itself costs almost nothing — and the answer you get is worth considerably more than any vendor’s benchmark table.