On coding performance alone, a 763B model handed out under MIT drew level with the summit — DeepSeek V4.1-Flash-1017 edged out Opus 5 and GPT-5.6 Sol on the SWE benchmark
On October 17, DeepSeek released V4.1-Flash-1017, an MIT-licensed open-weight model. It scored 77.0 on DeepSWE v1.1, the coding benchmark, edging out Opus 5 and GPT-5.6 Sol on that single point. You can self-host it or use a cheap API, but it still trails on general reasoning like HLE — and "silentl
On October 17, China's DeepSeek released the open-weight model "DeepSeek-V4.1-Flash-1017" on Hugging Face. The license is MIT, and it is positioned as a "minor update" to the previous V4.1-Flash. Yet it scored 77.0 on DeepSWE v1.1, a benchmark that measures software-development ability, edging out the closed frontier models Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0) on this single metric. The fact that anyone can download the weights and run the model locally or on-premise is the biggest talking point for developers.
A "minor update" overtook the frontier on a single coding metric
The released model has roughly 763B total parameters, a context window of up to 1 million tokens, and a maximum of 384,000 tokens per output. DeepSeek says it improved on the previous version by 2.0 to 11.1 points across all 11 benchmarks it published. Here are the key numbers.
| Benchmark | What it measures | V4.1-Flash-1017 | Previous version | Comparison (closed) |
|---|---|---|---|---|
| DeepSWE v1.1 | Software development | 77.0 | 74.2 | Opus 5: 74.0 / GPT-5.6 Sol: 73.0 |
| Terminal-Bench 4.0 | Terminal operations | 42.3 | 31.2 | Opus 5 reported to lead |
| Apex | Compound reasoning | 72.1 | — | — |
| HLE | Hard-problem reasoning | 42.6 | — | Opus 5: 56.3 |
DeepSeek also explains that this update was tuned to cut task execution time and output token volume, finishing "the same job faster and cheaper." For use cases like coding agents that call tools dozens of times, shorter outputs keep latency and billing from piling up — a quiet but real benefit here.
Strong on SWE, but still a gap in general reasoning
That said, "beating Opus 5" holds only on DeepSWE, a coding metric. On HLE, which measures hard-problem reasoning, Opus 5 scores 56.3, far ahead of this model's 42.6, and on Terminal-Bench 4.0 Opus 5 is reported to lead as well. Its strengths clearly lean toward software development, and it is not at the stage of replacing the frontier across every task. If you are considering adoption, the realistic approach is to evaluate it narrowly on the benchmarks closest to your own use case.
Sidestepping the "open or closed" binary — handed out under MIT, with API prices held steady
The biggest significance of this model lies less in the scores themselves than in the fact that the weights come down under MIT. Even organizations that cannot send code outside can load it onto their own GPU servers or private cloud and build coding assistance in-house, avoiding dependence on closed APIs and cross-border data flows. At 763B, the scale is heavy for a personal PC and presupposes quantization or multiple GPUs, but it is within reach for an on-premise server. The pricing for the accompanying API is as follows.
| Item | Price (standard / off-peak) |
|---|---|
| Input (cache miss) | $0.15 / 1M tokens |
| Input (cache hit) | $0.003 / 1M tokens |
| Output | $0.60 / 1M tokens |
| Search / Fetch API | $0.75 per 1,000 successful calls (free until October 31) |
During weekday peak hours, the per-token price roughly doubles. The Search / Fetch API has a cap of 5 requests per second per account. Being able to pursue both "run it yourself" and "use a cheap API" widens a set of options that used to be closed-only.
Same name, swapped-out internals — the blind spot of a "silent deploy"
According to reports, this update was swapped in under the existing API identifier, with no changelog. In other words, existing users calling the API were already receiving the behavioral changes before the documentation was in place. This is one example of the trend toward treating an AI API not as a "fixed-version dependency" but as a "pipeline whose internals change continuously."
Behind the convenience, such silent updates have also been flagged as a reproducibility risk. If output quirks or token volumes change silently, evaluation results, cost estimates, and downstream automation can quietly drift out of alignment. The countermeasures are simple.
- Pin versions where you can. Pin a dated tag or the open-weight release to keep a "known-working state" at hand.
- Run regression evaluations on critical paths with every update. Monitor both quality and cost with a small test set.
- Keep the open weights archived. Even if the cloud-side behavior changes, keep an escape route that lets you reproduce and verify with the same weights.
The more cases pile up of open-weight models drawing level with the frontier on a specific benchmark, the more options beyond "standardizing on the single strongest closed model" land in developers' hands. V4.1-Flash-1017, at least on the single point of coding, is one example showing that this option has come within practical reach. That said, it hands over homework at the same time: discerning its strengths, and managing versions.
References: DeepSeek official news / MarkTechPost: Scores 77.0 on DeepSWE / AI Weekly: Opens weights and $0.75 Search API / SmartScope: Benchmarks & price / Hugging Face model card
