DeepSeek V4.1 Flash hands-on tests and comparisons

Agent Arena ranks DeepSeek-V4.1-Flash-20261017 #11 overall and #1 among open models, with +5.1% net gain over 7.5K sessions and $0.20 median cost per task, about 27% of GPT-6.1 Sol's cost for 37% of its gain.

The author highlights V4.1 Flash 1017's score of 45 on a chart and notes OpenCode now offers the 1017 snapshot free (including Go) with a 1.05M context window and up to 393K output tokens.

Artificial Analysis gives DeepSeek V4.1 Flash 1017 an Intelligence Index of 45, six points above the previous V4.1 Flash and tied with GLM-5.3 for second among open-weights models.

Per DeepSeek's own table, V4.1 Flash scores 74.2 vs Opus 5's 74.0 on DeepSWE but 36.8 vs 56.3 on Humanity's Last Exam, so coding has caught up while hard reasoning still trails by about 20 points.

Mia-AiLab's 2.9 bpw EXL3 EAQ v1 quant of DeepSeek V4.1 Flash 1017 reaches a 67.5 tok/s median on 2× DGX Spark with TP2 and MTP on, and 72.5 tok/s at temp 1.0.

On the same corpus and prompt, DeepSeek 1017 scored 45/49 recall vs Kimi K3's 46/49 with zero hallucinations for both; 1017 won the relational check 5/5 and was 2.1x faster and 4.4x cheaper on a 162k-token prompt, becoming the new default compaction model.

Across four open-weights releases on one coding benchmark, DeepSeek V4.1 1017 came out strongest but inconsistent, not clearly better than the prior 4.1 on average and worse in some environments, though far cheaper to run.

A mini voxel reef was generated via DeepSeek Harness with the V4.1 Flash 1017 online API, with same-prompt local and cross-model tests planned.

DeepSeek released an updated V4.1-Flash with open weights and opened its Search API the same day; its own table shows DeepSWE 77.0 vs Opus 5's 74.0 but wins only 4 of 11 items, at about ¥32 per coding task and ¥120 per 1,000 searches.

Citing FundaAI, the post argues 1017's DeepSWE score and per-task cost do not mean it beats GPT-6; it sits with Kimi K3 and MiMo-V2.6 Pro at the open-weight top, not closer to closed leaders.

EXL3 2.9bpw quantization plus TensorFold hits 56.8 tok/s text and 92.2 tok/s code decoding, with 1M-token prefill at 1,246 tok/s passing needle retrieval.

A news digest notes Together AI now serves DeepSeek-V4.1-Flash-1017 at the same price, plus user reports of $2 per 100M tokens, beating Kimi K3 on hand tasks, and a 13-issue Issue-to-PR test.