DeepSeek-V4.1-Flash · 官方评测披露
DeepSeek-V4.1-Flash 当前收录 22 项测试的官方记录。已完成列明材料的核查。
模型发布 · DeepSeek V4.1 Flash
官方材料与核查范围
All listed official model-card, official News/API, metadata and directly linked technical-report materials for this release were read through the requested scope. Tables without an explicit unit retain an empty unit; figures without exact labels were not estimated. 核对时间 。
当前资料不足以还原首发。 No reliable historical snapshot proving the complete launch-day material set was obtained. Official publication dates and current materials do not substitute for a launch snapshot.
- DeepSeek-V4.1-Flash official Hugging Face model card · 主文 · HF revision 314113f8a7dadb388215396ba5fbee2c610a04d6; current card snapshot公布 未确认 · 采集 · 已核查 Official model-card Markdown read in full; benchmark rows, notes, configuration statements and linked reports checked.
- DeepSeek-V4.1-Flash official Hugging Face model metadata · 主文 · author=deepseek-ai; repo createdAt=2026-09-10T02:17:58.000Z (repository upload time only); lastModified=2026-10-16T09:27:28.000Z; sha=314113f8a7dadb388215396ba5fbee2c610a04d6公布 未确认 · 采集 · 已核查 Official model API metadata checked for author, repository identity, revision and timestamps; createdAt was never used as release date.
- DeepSeek V4.1 Flash:更强、更快、更普惠 · 主文 · Official DeepSeek News index entry dated 2026-09-10; index downloaded and retained locally公布 · 采集 · 已核查 The official dated entry or current service table was read for this release scope; aliases were not treated as weight checkpoints.
- DeepSeek-V4.1-Flash Technical Report · 附件 · Official V4.1-Flash report: 51/51 PDF pages read; Table 1, Table 3, Table 4, Table 5, multi-agent results and exact-label results were checked. CyberGym, SEC-Bench Pro and ExploitGym were retained as local audit only and excluded from regular mappings.公布 未确认 · 采集 · 已核查 Official V4.1-Flash report: 51/51 PDF pages read; Table 1, Table 3, Table 4, Table 5, multi-agent results and exact-label results were checked. CyberGym, SEC-Bench Pro and ExploitGym were retained as local audit only and excluded from regular mappings.
配置
- DeepSeek-V4.1-Flash reasoning_effort=100 (Max display) · reasoning_effort=100 exactly as reported; temperature=1.0, top_p=0.95. All Instruct results use the maximum effort setting. 100未知项 Display label Max maps to the original numeric reasoning_effort=100; it does not replace the raw setting.
- Codex v0.147.0 · Codex v0.147.0 standard app-server mode with adapted tool schemas. Max reasoning effort; Linux containers, temperature=1.0, top-p=0.95, 1M context and max_steps=500 as stated in the report note. This is an ordinary single-agent public scaffold run, not the Section 5.3.5 multi-agent experiment. Max未知项 Public reproducible ordinary single-agent scaffold at explicit Max reasoning effort. Presentation uses the existing max mode; scaffold and version remain conditions and do not create a separate effort tier.
- OpenCode v1.18.15 · OpenCode v1.18.15 build agent with shell and file tools and native task delegation. Max reasoning effort; Linux containers, temperature=1.0, top-p=0.95, 1M context and max_steps=500 as stated in the report note. This is an ordinary single-agent public scaffold run, not the Section 5.3.5 multi-agent experiment. Max未知项 Public reproducible ordinary single-agent scaffold at explicit Max reasoning effort. Presentation uses the existing max mode; scaffold and version remain conditions and do not create a separate effort tier.
- Pi v0.84.2 · Pi v0.84.2 RPC mode with file and shell tools plus search extension. Max reasoning effort; Linux containers, temperature=1.0, top-p=0.95, 1M context and max_steps=500 as stated in the report note. This is an ordinary single-agent public scaffold run, not the Section 5.3.5 multi-agent experiment. Max未知项 Public reproducible ordinary single-agent scaffold at explicit Max reasoning effort. Presentation uses the existing max mode; scaffold and version remain conditions and do not create a separate effort tier.
- mini-SWE · mini_swe_v2 port with a single bash tool and submission marker. Max reasoning effort; Linux containers, temperature=1.0, top-p=0.95, 1M context and max_steps=500 as stated in the report note. This is an ordinary single-agent public scaffold run, not the Section 5.3.5 multi-agent experiment. Max未知项 Public reproducible ordinary single-agent scaffold at explicit Max reasoning effort. Presentation uses the existing max mode; scaffold and version remain conditions and do not create a separate effort tier.
- DeepSeek Harness Standard · DeepSeek Harness Standard with the full SDK profile and 26 initial tools. Max reasoning effort; Linux containers, temperature=1.0, top-p=0.95, 1M context and max_steps=500 as stated in the report note. This is an ordinary single-agent public scaffold run, not the Section 5.3.5 multi-agent experiment. Max未知项 Public reproducible ordinary single-agent scaffold at explicit Max reasoning effort. Presentation uses the existing max mode; scaffold and version remain conditions and do not create a separate effort tier.
- DeepSeek Harness PTC · DeepSeek Harness PTC with run_code and 24 underlying tools. Max reasoning effort; Linux containers, temperature=1.0, top-p=0.95, 1M context and max_steps=500 as stated in the report note. This is an ordinary single-agent public scaffold run, not the Section 5.3.5 multi-agent experiment. Max未知项 Public reproducible ordinary single-agent scaffold at explicit Max reasoning effort. Presentation uses the existing max mode; scaffold and version remain conditions and do not create a separate effort tier.
- Claude Code v2.1.105 · Claude Code v2.1.105 native tool interface. Max reasoning effort; same scaffold and benchmark settings as Table 4. This is an ordinary single-agent public scaffold run, not the Section 5.3.5 multi-agent experiment. Max未知项 Public reproducible ordinary single-agent scaffold at explicit Max reasoning effort. Presentation uses the existing max mode; scaffold and version remain conditions and do not create a separate effort tier.
- Claude Code v2.1.238 · Claude Code v2.1.238 native tool interface. Max reasoning effort; same scaffold and benchmark settings as Table 4. This is an ordinary single-agent public scaffold run, not the Section 5.3.5 multi-agent experiment. Max未知项 Public reproducible ordinary single-agent scaffold at explicit Max reasoning effort. Presentation uses the existing max mode; scaffold and version remain conditions and do not create a separate effort tier.
- Claude Code v2.1.251 · Claude Code v2.1.251 native tool interface. Max reasoning effort; same scaffold and benchmark settings as Table 4. This is an ordinary single-agent public scaffold run, not the Section 5.3.5 multi-agent experiment. Max未知项 Public reproducible ordinary single-agent scaffold at explicit Max reasoning effort. Presentation uses the existing max mode; scaffold and version remain conditions and do not create a separate effort tier.
- Claude Code v2.1.259 · Claude Code v2.1.259 native tool interface. Max reasoning effort; same scaffold and benchmark settings as Table 4. This is an ordinary single-agent public scaffold run, not the Section 5.3.5 multi-agent experiment. Max未知项 Public reproducible ordinary single-agent scaffold at explicit Max reasoning effort. Presentation uses the existing max mode; scaffold and version remain conditions and do not create a separate effort tier.
- Claude Code four-version average · Average computed from the unrounded Pass@1 scores of four Claude Code versions. Max reasoning effort; same scaffold and benchmark settings as Table 4. Max未知项 Claude Code version-specific results are retained raw and are not mapped as ordinary effort tiers.
- DeepSeek-V4.1-Flash multi-agent 8h · Official V4.1 report Section 5.3.5; up to three rollouts per task and an 8-hour per-rollout deadline. 未知项 Multi-agent experiments are explicitly excluded from the regular presentation under SPEC.
- DeepSeek-V4.1-Flash single-agent 8h · Official V4.1 report Section 5.3.5; strongest available single-agent baseline at the 8-hour deadline. 未知项 Multi-agent experiments are explicitly excluded from the regular presentation under SPEC.
- DeepSeek-V4.1-Flash multi-agent 20h · Official V4.1 report Section 5.3.5; no-GPU subset and 20-hour per-rollout deadline. 未知项 Multi-agent experiments are explicitly excluded from the regular presentation under SPEC.
- DeepSeek-V4.1-Flash single-agent 20h · Official V4.1 report Section 5.3.5; strongest available single-agent baseline at the 20-hour deadline. 未知项 Multi-agent experiments are explicitly excluded from the regular presentation under SPEC.
- DeepSeek Harness Minimal · DeepSeek Harness Minimal with a single bash tool (Appendix B.1). Max reasoning effort; Linux containers, temperature=1.0, top-p=0.95, 1M context and max_steps=500 as stated in the report note. This is an ordinary single-agent public scaffold run, not the Section 5.3.5 multi-agent experiment. Max未知项 Public reproducible ordinary single-agent scaffold at explicit Max reasoning effort. Presentation uses the existing max mode; scaffold and version remain conditions and do not create a separate effort tier.
直接前代与后续发布
当前收录未确认直接前代。 官方关系依据
测试记录 · 22
- GPQA Diamond · Diamond subset; task snapshot unspecified本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
- Humanity's Last Exam · Full HLE set as distinguished from the card-marked text-only subset; task revision unspecified本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
- Codeforces · Rating; contest snapshot unspecified本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
- MathArena Apex · Version unspecified; Pass@1 as printed in V4.1-Flash card本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
- Terminal-Bench · 2.1; Pass@1 as printed in V4.1-Flash card本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
- Terminal-Bench · 3.0; Pass@1 as printed in V4.1-Flash card本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
- Terminal-Bench · 4.0; Pass@1 as printed in V4.1-Flash card本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
- DeepSWE · v1.1; Resolved metric as printed in V4.1-Flash card本次发布有 13 条成绩记录,含 13 条精确值(可含历史记录)
- ProgramBench · Metric Almost@1 as printed in V4.1-Flash card; task version unspecified本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
- NL2Repo-Bench · Version unspecified; Score metric as printed in V4.1-Flash card本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
- AutomationBench · v1.0.6; Pass@1 as printed in V4.1-Flash card本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
- Agents' Last Exam · Task revision unspecified; Pass@1 as printed in V4.1-Flash card本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
- Humanity's Last Exam · With-tools evaluation; task subset unspecified本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
- Chartography · With tools; task revision unspecified本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
- BabyVision · With tools; task revision unspecified本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
- ZeroBench-main · With tools; Pass@5 as printed in V4.1-Flash card本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
- Humanity's Last Exam · Text-only subset marked † in DeepSeek V4.1-Flash card; task revision unspecified本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
- Terminal-Bench · 2.1 (Pass@1); across agent scaffolds at Max reasoning effort; unit not printed本次发布有 11 条成绩记录,含 11 条精确值(可含历史记录)
- ProgramBench · High-confidence 172-task subset; Almost@1; multi-agent; 8-hour deadline本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
- ProgramBench · High-confidence 172-task subset; Almost@1; single-agent; 8-hour deadline本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
- FrontierSWE v2 · No-GPU public subset; Mean@5; multi-agent; 20-hour deadline本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
- FrontierSWE v2 · No-GPU public subset; Mean@5; single-agent; 20-hour deadline本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)