DeepSeek ships V4 Flash as open weights, and the agentic-coding gap narrows

On July 31, 2026, DeepSeek published the production build of V4 Flash, tagged 0731, as MIT-licensed open weights. The interesting part is not another cheap model. It is where the gains landed: DeepSeek’s own numbers and one credible outside evaluation both point at agentic coding and tool use, the exact territory where proprietary frontier models have held a clear lead.

What changed from the April preview

DeepSeek describes 0731 as a re-post-trained version of the April V4 Flash preview rather than a new architecture. The model is a Mixture-of-Experts design of roughly 304 billion total parameters that activates only a small fraction per token, and DeepSeek reports it now beats the much larger V4-Pro (Preview) across the agentic benchmarks it published.

The jumps DeepSeek reports are large. On Terminal Bench 2.1, a test of multi-step work in a real shell, the model card lists 82.7 for 0731 against 72.1 for the preview. On DeepSWE, an agentic software-engineering benchmark, it lists 54.4 against the preview’s 12.8. It also reports 76.7 on Cybergym and 70.3 on Toolathlon-Verified, and it adds a three-level reasoning_effort control so developers can trade latency and cost against deliberation. These are the vendor’s own figures, so they set expectations rather than settle them, but the shape is consistent: the work went into agents and coding, not leaderboard reasoning scores.

The independent read matters more than the self-report

Vendor benchmarks are cheap to publish and easy to tune toward. Third-party evaluation is the part that carries weight, and here there is one. ARC Prize, which runs the ARC-AGI reasoning benchmarks and publishes verified per-model results, tested 0731 at maximum effort and recorded 89.0 percent on ARC-AGI-1 at about $0.02 per task and 61.4 percent on ARC-AGI-2 at about $0.04 per task. ARC-AGI-2 is deliberately hard and resistant to memorization, so a 61.4 percent result at four cents a task is a genuine price-performance data point rather than a marketing number.

Put the two together and the claim gets more grounded than a launch post alone would support: an open-weight model is posting frontier-adjacent agentic and reasoning results at a cost structure the proprietary leaders cannot match on price. DeepSeek has made that argument before on raw reasoning. Doing it on agentic coding is the newer, more consequential move.

What it means for developers

For working teams, MIT-licensed weights change the option set, not just the price. The model can be self-hosted, fine-tuned, or run in sovereign and air-gapped environments where sending code to a proprietary API is a non-starter, and DeepSeek’s hosted endpoint remains far cheaper per token than the proprietary frontier. On the integration side, DeepSeek adapted the endpoint for the Responses API format and for Codex-style agent workflows, which lowers the switching cost for anyone already wired into an agentic coding harness.

The honest framing is a floor, not a coronation. A cheap open model that lands within range of proprietary agents on published benchmarks resets what teams should expect to pay for competent coding automation, and it puts pressure on every vendor pricing a mid-tier coding model. Whether it holds up on messy real-world repositories, rather than curated benchmark tasks, is the question a single week of results cannot answer.

Caveats and what to watch

Two caveats belong on any procurement conversation. First, benchmark strength and production reliability are different properties, and the gap between them is where agentic models tend to disappoint. Second, DeepSeek’s China origin remains a real constraint for some organizations regardless of the MIT license, which is why the self-host path matters as much as the hosted API.

Stackmaven’s catalog still carries DeepSeek V4 with a “watch” label. The signals that would move it: independent coding-agent results outside the vendor’s own benchmark set, evidence that the small active-parameter count holds quality on long tool-use chains, and adoption in serious open-source agent stacks. Stackmaven will reassess on or around November 6.

Sources cited

1. DeepSeek-V4-Flash-0731 model card (DeepSeek, Hugging Face) huggingface.co
2. DeepSeek V4 Flash 0731 verified results (ARC Prize) arcprize.org
3. DeepSeek upgrades V4-Flash-0731 with major agentic and coding gains (MarkTechPost) www.marktechpost.com