The DeepSeek V4.1 Flash open weights coding agent has been released with a download package on Hugging Face, scoring 77.0 percent on the DeepSWE benchmark for software engineering tasks. The result positions the model ahead of Anthropic’s Claude Opus 5 on the same evaluation, while Anthropic has not yet published an open weights response. DeepSeek published the new build alongside technical notes describing how the coding agent handles multi-file refactors and repository-scale reasoning.
DeepSeek V4.1 Flash open weights coding agent benchmark results
According to the release notes shared on Hugging Face, the DeepSeek V4.1 Flash open weights coding agent reached a 77.0 percent pass rate on DeepSWE, a benchmark that measures a model’s ability to solve real GitHub issues across multiple programming languages. The same notes cite a Claude Opus 5 result of 74.2 percent on the same test, although the comparison has not been independently verified on a fresh leaderboard refresh. DeepSeek’s evaluation scripts were also released so that outside researchers can attempt to replicate the figure on identical hardware.
Beyond the headline DeepSWE number, DeepSeek reported scores on SWE-bench Verified and a smaller internal repository-navigation suite. SWE-bench Verified landed within two points of the previous Flash generation, while the internal suite reportedly improved by roughly four points. The company framed the gains as a function of better tool-use formatting and an expanded context window for repository-wide retrieval rather than a wholesale architectural rewrite.
DeepSeek V4.1 Flash open weights coding agent pricing and deployment
DeepSeek listed the hosted price for the new model at $0.15 per million input tokens, with a higher rate for output tokens that varies by region. That figure, if confirmed by independent testers, would undercut several frontier coding assistants currently sold to enterprise developers, including the more expensive tiers of Anthropic’s Claude product line. Independent hosting is also possible because the weights are downloadable, which removes the need to route every request through a DeepSeek-controlled endpoint.
Operators running local inference can load the model with vLLM or the standard Hugging Face transformers pipeline, according to the model card. The card recommends at least 80 gigabytes of VRAM for full-precision inference, while quantized variants in the GGUF format can fit on a single high-end consumer GPU. DeepSeek did not publish latency benchmarks, leaving throughput comparisons to community evaluators who have begun publishing first impressions on X and various Discord servers.
DeepSeek V4.1 Flash open weights coding agent compared with Gemini 4 Argon
The same week that DeepSeek pushed its weights, Google shipped Gemini 4 Argon, a closed-source release that internal reviewers described as focused on agentic code generation and longer-horizon planning. Google has not yet released a public DeepSWE score for Argon, so a head-to-head comparison rests on third-party harnesses and qualitative impressions rather than matched numbers. Developers who have tested early access builds report that Argon performs well on tool orchestration but still trails OpenAI’s top offering on certain frontend tasks, a gap Gemini 4 attempts to close through a redesigned chain-of-thought routine.
For organizations evaluating the two, the dividing line is licensing and hosting rather than raw accuracy. DeepSeek offers a downloadable artifact that can be audited and self-hosted, while Argon runs only through Google’s own API and Vertex AI endpoint. Procurement teams have already begun asking compliance officers how an open weights model from a Chinese lab fits into existing data residency policies, a question that may shape enterprise adoption as much as benchmark scores.
DeepSeek V4.1 Flash open weights coding agent limits and oversight
DeepSeek’s release notes include a model card that lists known limitations, including weaker performance on certain Rust and Kotlin tasks and occasional refusals on dual-use security questions. The card does not describe what safety fine-tuning was performed beyond standard reinforcement learning from human feedback, which leaves open questions for researchers who want to compare alignment behavior across coding models. Several academic groups have already requested access to the training data summary, and DeepSeek has indicated that a more detailed system card will follow.
Outside observers caution that DeepSWE, like SWE-bench before it, can be partially gamed by models that have seen similar tasks during pretraining. Independent replication of the 77.0 percent result will be necessary before the figure can be cited in procurement documents. For now, the new DeepSeek V4.1 Flash open weights coding agent release marks another data point in a market where open weights vendors are increasingly matching closed-source labs on narrowly scoped evaluations.
Released on October 17, 2026 as the V4.1-Flash-1017 checkpoint, the DeepSeek V4.1 Flash open weights coding agent also registered a 42.6 score on the Humanity’s Last Exam reasoning suite, a strong showing that still leaves Claude Opus 5 in front at 56.3 on that benchmark. On DeepSWE, DeepSeek reported GPT-5.6 Sol at 73.0, reinforcing the narrative that the Flash refresh has leapfrogged closed-source coding rivals on software engineering tasks while ceding ground on broader reasoning evaluations. Bloomberg’s same-day coverage described the release as a direct attempt by DeepSeek to pressure OpenAI and Anthropic on price-performance.
Pricing details sharpened in the days after launch: the DeepSeek serve tier lists $0.15 per million input tokens and $0.60 per million output tokens, with off-peak traffic charged at half those rates. The bundled Search and Fetch tool APIs run at $0.75 per 1,000 successful calls, though server-side tools remain free for hosted customers until November 1, a promotional window aimed at seeding developer integrations. Existing V4.1 customers see their API endpoint auto-upgrade to the Flash variant, so no code changes are required to start routing requests through the refreshed weights.
Compared with Gemini 4 Argon’s introductory rates of roughly $2 input and $10 output per million tokens, the DeepSeek V4.1 Flash open weights coding agent effectively prices frontier coding assistance at less than a tenth of Google’s headline figure. Analysts tracking the segment argue that this gap, combined with downloadable weights, gives the Flash release credible claim to the title of cheapest serious coding agent on the market heading into late 2026.
Source: https://dev.to/paullovelace/two-ai-giants-shipped-this-week-gemini-4-argon-vs-deepseek-v41-flash-12ac