DeepSeek V4 Flash 0731: The 13B-Active Agent Workhorse That Outranks Its Own Flagship

On July 31, 2026, DeepSeek moved V4 Flash from preview into official public beta as DeepSeek-V4-Flash-0731, and in doing so quietly upended the usual ordering of a model family launch: this is the case of a small model beating its own flagship. The 0731 checkpoint keeps the exact architecture of the April preview — a sparse 284B-total / 13B-active Mixture-of-Experts backbone — and was only re-post-trained, yet DeepSeek's own changelog claims it now outperforms the far larger V4-Pro preview (1.6 trillion total parameters, 49B active) on essentially every agentic benchmark the company publishes. That a 284B open-weight model can outscore a 1.6T flagship on long-horizon, tool-calling, terminal-driving tasks at a fraction of the price is not just a spec-sheet curiosity; it is the clearest signal yet that the frontier of agentic performance is shifting from raw parameter count toward post-training quality. This article walks through what actually shipped, the benchmark table with the vendor-claim caveat it deserves, how the model compares against its own family and the strongest proprietary rivals, and what the pricing and deployment story means for developers building agent fleets today.

What Shipped: Same Architecture, Rewritten Post-Training

The headline detail of the 0731 release is that nothing about the model itself changed. DeepSeek's API changelog states it plainly: DeepSeek-V4-Flash-0731 keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained. That is a narrow, precise announcement — and it matters because it tells you where the performance jump allegedly comes from.

The update applies only to the deepseek-v4-flash API model. DeepSeek's own app, website, and the V4-Pro API were untouched, and the changelog confirms the official (non-preview) V4-Pro release is still pending with no firm date. In other words, the company productionized its smaller, cheaper agent model before its flagship — and marketed the small model's agent scores as far exceeding the big one's. That sequencing is the real story of the day: DeepSeek is positioning V4 Flash, not V4 Pro, as the agent workhorse that carries production traffic.

The DSpark Connection

One structural detail is easy to miss: the 0731 checkpoint shares its backbone with DeepSeek-V4-Flash-DSpark, which means it ships with a speculative decoding module. DeepSeek calls the mechanism DSpark, and it is enabled in vLLM with a single flag:

--speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'

SGLang users flip it on with --speculative-algorithm DSPARK, where draft and target weights come from the same checkpoint. This is a meaningful deployment detail: it is how a 284B model reaches high per-token throughput on a modest node rather than a multi-GPU rack.

The Agent Benchmarks, Read With the Required Caveat

Before diving into scores, the honesty requirement: every number in the table below comes from DeepSeek's own changelog, produced on DeepSeek's own evaluation harness, run in minimal (agent-framework) mode at maximum reasoning effort. No independent lab had reproduced any of these figures as of release day, and much of the secondary coverage repeated them without that caveat. Treat the suite as vendor claim awaiting third-party verification — which does not make it meaningless, only provisional. The one semi-independent data point comes from Artificial Analysis, which scored the reasoning build (max effort) around 50 on its Intelligence Index; but that measures general intelligence, not this agent suite specifically.

Two internal DeepSeek sets, DSBench-FullStack and DSBench-Hard, are not public, which matters for any head-to-head reading against models evaluated only on public tasks. But the pattern holds across the public sets: Flash 0731 sits at or near the top among open-weight models, ahead of its own family, and within striking distance of the strongest proprietary systems.

What Changed From Preview

The single biggest move is on DeepSWE, which vaults from 7.3 (preview) to 54.4 — a roughly seven-fold jump that alone validates the post-training investment. Cybergym more than doubles from 38.7 to 76.7, AutomationBench jumps from 10.8 to 25.1, and Terminal Bench climbs from 61.8 to 82.7. When a re-post-train-only release claims agent gains of this magnitude, the launch is effectively an argument that the remaining frontier in agent performance lives in post-training, not scale. If the numbers hold under third-party evaluation, that argument gets much stronger; if they do not, the release still stands on price-performance economics that do not depend on the benchmarks at all.

How It Compares to the Field

The comparison readers care about most is the intra-family one — Flash 0731 versus the V4-Pro preview — because it is so counterintuitive. DeepSeek's own numbers put the 284B Flash ahead of the 1.6T Pro on essentially every published agent suite despite activating about 13B parameters per token (roughly a quarter of Pro's 49B). The release is practically a statement of intent: DeepSeek is betting that agent quality is driven by high-quality, tool-use-heavy post-training, not by parameter count.

Against the broader frontier, Flash 0731 lands just ahead of GLM-5.2 (82.7 vs 81.0 on Terminal Bench; 54.4 vs 46.2 on DeepSWE) and runs a tight race with Opus 4.8, which edges it out on most rows but at a far higher price tier. On Agents' Last Exam the gap narrows to a rounding error (25.2 vs 25.7). The ARC-AGI suite — a useful independent check — shows Flash at max effort posting 89 on ARC-AGI-1 Semi-Private and 61.4 on ARC-AGI-2 Semi-Private, flagship-adjacent reasoning at a fraction of the cost.

Where the Open-Weight Race Stands

The timing puts Flash 0731 in interesting company. Kimi K3 — the 3T open-weight frontier model we covered at launch — posted 76.1 on Terminal Bench; Flash 0731's 82.7 clears it on a much smaller active parameter footprint. Open-weight models now own long-horizon terminal work, and DeepSeek is pushing its cost model down the hardest with this 284B base.

Pricing, Speed, and the Economics of Agent Fleets

The operational math is where this release gets interesting. On OpenRouter, the model lists at $0.065 per 1M input tokens and $0.14 per 1M output, with the highest-throughput providers marking it up toward $0.14 / $0.28 and cache-hit input landing near $0.02. DeepSeek's own first-party rate sits at the higher end, $0.14 cache-miss input / $0.28 output with a cache-hit input near $0.002 — a deep discount when the same context is replayed. Whichever route you use, the context window is 1M tokens, with long agent traces and whole repositories fitting without aggressive compaction.

One caveat: verbosity. This is a verbose reasoning model. Artificial Analysis noted it generates far more output tokens than the median across its eval suite, which erodes some of the per-token price advantage. Budget on tokens generated, not just price per token. Even so, it remains one of the cheapest frontier-adjacent options available, and it also supports 2,500 concurrent requests versus Pro's 500 — a five-fold headroom gap that matters more to production agent fleets than any single benchmark score.

The reasoning_effort Dial

The 0731 build adds a control layer: reasoning_effort with three levels, low, high, and max, governing how much deliberation the model spends before answering. The practical guidance emerging from DeepSeek's harness and from deployments in tools like OpenCode is that max suits multi-step agent loops (plan, run tools, self-correct), while high is faster and cheaper for single-file edits and review passes. Given the verbosity data above, dropping to high when a task does not need long-horizon planning is the easiest cost lever you can pull.

How to Run It Today

API Path

The simplest route is the official deepseek-v4-flash API model, now in public beta on OpenRouter and DeepSeek's first-party API. Note the alias migration: DeepSeek retired the deepseek-chat and deepseek-reasoner API aliases on July 24, 2026 — any integration still calling the old names is already broken and must switch to the deepseek-v4-flash / deepseek-v4-pro model IDs to receive the 0731 upgrade at all.

Self-Hosted Path

For self-hosting, the vLLM command serves the checkpoint on a single 4xGB300 node with fp8 KV cache and DSpark speculative decoding:

vllm serve deepseek-ai/DeepSeek-V4-Flash-0731 \
  --trust-remote-code --kv-cache-dtype fp8 --block-size 256 \
  --data-parallel-size 4 --enable-expert-parallel \
  --moe-backend deep_gemm_mega_moe \
  --attention-config '{"use_fp4_indexer_cache": true}' \
  --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'

There is no Jinja-format chat template in this release; instead the model ships an encoding folder with Python scripts for turning OpenAI-format messages into the model input string and back. SGLang users enable the same speculative path with --speculative-algorithm DSPARK without pointing at a separate draft model, since target and draft weights come from the same checkpoint.

A Squeeze on Frontier Pricing

Zooming out, 0731 is a data point in a squeeze now visibly reshaping frontier-lab economics. An MIT-licensed 284B model at roughly $0.07 / $0.14 per million tokens on OpenRouter posting a Terminal Bench score above many proprietary rivals sets a floor: any lab charging real money for agentic coding now has to explain what the premium buys. OpenAI cut one of its production models by 80% shortly before, in part by a model rewriting its own serving kernels to reduce serving cost — a compounding loop that is the only way a closed lab keeps pace with a competitor that gives the weights away. On the other side, Anthropic's top-tier agent tier prices land at a huge multiple on cached input, a gap that only holds if the quality difference sits in the part of your workload that pays for it. On extraction, classification, and the terminal-loop agent work this release targets, it increasingly does not.

Open Questions Before You Bet Production On It

Finally, the honest caveats. As of release day, no 0731 weights had appeared on the public MIT-licensed Hugging Face repository — it still hosted the April preview — so open-weight status for this checkpoint is an open question, not a settled fact. The announced peak/off-peak dynamic pricing on DeepSeek's first-party was likewise not yet in effect, pending an official date. And most importantly for engineers, the benchmarking gap: until a third party reproduces even one of these agent scores, the smart posture is to treat the chart as a promise to validate, not a fact to plan around. Run a side-by-side on your own production tasks before re-architecting your fleet.

For more of how these models shake out, see our deep dives on Kimi K3's Terminal Bench run, the Thinking Machines Inkling open-weight release, and how frontier labs are pricing close-to-frontier quality down. And if you are weighing whether to route agent work to a cheap open-weight model in the first place, the agentic risk conversation is the other half of that decision.