DeepSeek has released DeepSeek-V4.1-Flash-1017, a new model checkpoint that scores 77.0 on the DeepSWE benchmark and has opened its Search and Fetch APIs to developers. The update is available on the DeepSeek website, mobile app, and API.
The company describes this as a Pareto improvement, offering better performance and lower costs than the September V4.1-Flash release. Developers now have access to a cheaper coding model paired with a first-party web layer via a single API key.
Model specifications and performance
The 1017 checkpoint uses a 552 billion parameter backbone and 196 billion Engram parameters. It activates 8 billion parameters during prefill and 16 billion during decode. The context window supports 1 million tokens, with an output capacity of up to 384,000 tokens.
DeepSeek serves this model as deepseek-flash on its API. The September V4.1-Flash weights, which total 763 billion parameters in FP8 format, are available under an MIT license on Hugging Face. Specific hardware requirements and the 1017 weights remain undisclosed.
In testing, the new checkpoint beats the previous V4.1-Flash on all 11 benchmarks DeepSeek reported, improving scores by between 2.0 and 11.1 points. It leads the field on DeepSWE v1.1 with a score of 77.0, surpassing Claude Opus 5, which scored 74.0.
However, the model trails on the Human Language Evaluation (HLE) benchmark without tools, scoring 42.6 compared to Claude Opus 5’s 56.3.
Architecture and training
DeepSeek-V4.1-Flash-1017 is an updated post-training checkpoint of the V4.1-Flash architecture. The technical details follow the V4.1-Flash report, which describes a multimodal Mixture-of-Experts model with 40 Transformer layers split into a 20-layer causal encoder and a 20-layer decoder.
The DeepSeek team states the model uses reinforcement learning with self-feedback for end-to-end project work, with examples ranging from game creation to 3D modeling. It is tuned for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. The company recommends using DeepSeek Harness v0.2.1-rc.1 or later.
The architecture includes several specific components designed to reduce compute:
- Causal Encoder-Decoder (CED): Projects the decoder global KV from final encoder hidden states, nearly halving prefill compute.
- Compressed Sparse Attention 2 (CSA2): Runs layers in Full, Reindex, or Reuse modes, sharing main KV and indexer keys across layers.
- FP4 main KV cache: Reduces the global KV to 890 bytes per token, roughly one-quarter of the V4-Flash size.
- SWA Bounded Replay: Shrinks the persistent KV cache to about one-eighth of the V4-Flash size.
- Reasoning effort: A scalar value from 1 to 100. API tiers map low, high, and max to 50, 75, and 100 respectively.
Benchmark results
The figures below are DeepSeek’s own data. The company ran Terminal-Bench 4.0 in Claude Code and DeepSWE in Mini-SWE-V2. ZeroBench-main uses Pass@5.
| Benchmark | V4.1-Flash-1017 | V4.1-Flash | Kimi K3 | GLM 5.3 | GPT 5.6-Sol | Claude Opus 5 |
|---|---|---|---|---|---|---|
| DeepSWE v1.1 | 77.0 | 74.2 | 67.5 | 66.9 | 73.0 | 74.0 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | 12.6 | 37.9 | 39.9 | 51.8 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | 46.7 | 48.8 | 45.8 | 50.3 |
| Agents’ Last Exam | 35.4 | 31.8 | 27.6 | 28.5 | 26.7 | 28.6 |
| Apex Bench v1.1 | 72.1 | 64.6 | 47.8 | 56.6 | 54.3 | 65.8 |
| ProgramBench | 29.1 | 20.3 | 17.5 | 19.0 | 23.0 | 37.0 |
| NL2Repo-Bench | 68.2 | 65.4 | 58.0 | 58.0 | 56.8 | 75.3 |
| HLE (no tools) | 42.6 / 44.0* | 36.8 / 39.1* | 43.5 | 42.0* | 44.5 | 56.3 |
| BabyVision (w/ tools) | 92.5 | 89.6 | 85.7 | Not disclosed | 88.9 | 94.1 |
*Text-only HLE subset.
The 1017 checkpoint leads on DeepSWE, Automation-Bench, Agents’ Last Exam, and Apex Bench. Claude Opus 5 retains the lead on Terminal-Bench 4.0, ProgramBench, NL2Repo, and HLE.
Comparison with open-weight rivals
| Feature | DeepSeek-V4.1-Flash-1017 | Kimi K3 | GLM-5.3 |
|---|---|---|---|
| Total parameters | 552B backbone + 196B Engram | 2.8T | 753B |
| Active parameters | 8B prefill / 16B decode | 104B | Not disclosed |
| Context | 1M | 1,048,576 | Not disclosed |
| License | MIT (V4.1-Flash); 1017 weights Not disclosed | Kimi K3 License | GLM-5.3 license |
| Modality | Text, image in; text out | Text, image | Text |
| DeepSWE v1.1 | 77.0 | 67.5 | 66.9 |
| Agents’ Last Exam | 35.4 | 28.3 | 28.5 |
| API price (1M input miss / output, peak) | $0.30 / $1.20 | Not disclosed | Not disclosed |
Search and Fetch APIs
The Search API returns ranked web results for a query. The Fetch API returns a single page as Markdown. DeepSeek states its in-house engine serves hundreds of millions of requests daily.
- Endpoints:
POST /v1/tools/searchandPOST /v1/tools/fetchonapi.deepseek.com. - Search options: Allows 1 to 50 results, optional page content, and operators like
site:,lang:,after:andbefore:. - Result fields: Each result includes a rerank score from 0 to 3 and an authority tier from 0 to 4.
- Price: CostsSource Read original →
