DeepSeek V4.1 Flash 1017 Tops Two HuggingFace Agent Leaderboards: 90.6% Terminal-Bench 2.1, 35.4 Agents-Last-Exam
The October 17 checkpoint of DeepSeek V4.1 Flash has been evaluated on two HuggingFace-hosted agentic benchmark leaderboards, and the scores show a meaningful lift over the base model.
On Terminal-Bench 2.1, the community-maintained HuggingFace leaderboard tracking 37 models, the 1017 checkpoint scores 90.6%. The base V4.1 Flash scored 82.7% on the same evaluation — a 7.9-point improvement. On Agents-Last-Exam, a separate HuggingFace benchmark with nine evaluated models, the 1017 checkpoint scores 35.4, up from 25.2 on the base model. Both checkpoints place first in their respective leaderboard fields as of the October 23 snapshot.
The October 17 checkpoint release added native multimodal input and a Search API alongside the post-training improvements reflected in the benchmark gains.
How This Fits the V4.1 Flash Profile
The 1017 checkpoint results add to a benchmark profile that already showed V4.1 Flash at the front of open-weight agentic evaluations:
- LiveBench 2.0 Agentic Coding: 77.3 — highest among all models on LiveBench’s September 2026 run, including closed frontier models
- LiveBench overall: 81.1
- DeepSWE: 77.0
- Cost per successful task on LiveBench: $0.029
The open-weight model’s $0.029 per LiveBench task compares to $6.904 for Claude Fable 5.1 — the top-scoring model on that leaderboard at 85.7 overall. V4.1 Flash scores 81.1 overall, 4.6 points lower, at roughly 1/240th the per-task cost.
Caveat: HuggingFace vs Official Leaderboards
Terminal-Bench 2.1 and Agents-Last-Exam are community-run HuggingFace leaderboards, not the official tbench.ai platform. Harness implementations and task sets may differ from commercial evaluations. The official tbench.ai Terminal-Bench 2.0 leaderboard places GPT-5.5 first at 84.7% using the NexAU-AHE scaffold — a different model set and version.
HuggingFace leaderboard submissions are vendor-submitted; independent validation is not guaranteed. The 37-entry Terminal-Bench 2.1 field and 9-entry Agents-Last-Exam field are both relatively thin compared to older benchmarks.
Pricing is unchanged from the base model. The 1017 checkpoint is available on HuggingFace.
Key Numbers
| Benchmark | 1017 Checkpoint | Base V4.1 Flash | Delta |
|---|---|---|---|
| Terminal-Bench 2.1 (HF) | 90.6% (#1/37) | 82.7% | +7.9 pts |
| Agents-Last-Exam (HF) | 35.4 (#1/9) | 25.2 | +10.2 pts |
| LiveBench Agentic Coding | 77.3 | — | — |
| DeepSWE | 77.0 | — | — |
| Cost/task (LiveBench) | $0.029 | — | — |