← October 19, 2026 briefing · added in the 20:00 KST update

DeepSeek V4.1-Flash-1017 follow-up: independent Terminal-Bench 4.0 scores of 34% and 31.8% trail the model card's 42.3% (community reports)

Verified (fact)

According to a Reddit post, the model card lists V4.1-Flash-1017's Terminal-Bench 4.0 score as 42.3%. On the same benchmark, Artificial Analysis (AA) measured 34% and the Vals Index 31.82%. The model card states the score was measured with the Claude Code harness. The poster speculated that the gap arises because the external evaluators used a different harness. For comparison, Fable 5 scored 42% on AA and 41.41% on Vals. A GeekNews post reported that the model scored 45 on the AA Intelligence Index, against a median of 19 for open-weight models of similar size.

Needs fact-check (claim)

Terminal-Bench 4.0 score of 42.3% per the model card (measured with the Claude Code harness).

Why it matters (interpretation)

This shows that agentic benchmark scores for the same model can differ by nearly 10 percentage points depending on the harness and measurement conditions. Rather than picking models from announced figures alone, they should be compared alongside independent evaluations. Even so, on external indices it still ranks near the top among open models of its size.

Confidence medium · official source pending

Sources

Story thread

  1. 2026-10-17 DeepSeek releases 'V4.1-Flash-1017' on Hugging Face; Terminal-Bench 4.0 rises from 31.2 to 42.3 (community reports)
  2. 2026-10-17 DeepSeek V4.1 Flash 1017 follow-up: company's first CED (Causal Encoder-Decoder) architecture, native image understanding; 552B total, 8B input/16B output active (community reports)
  3. 2026-10-18 DeepSeek V4.1-Flash-1017 follow-up: official release confirmed with MIT-licensed weights, rolled out to web, app and API at unchanged prices (report)
  4. 2026-10-19 DeepSeek V4.1-Flash-1017 follow-up: independent Terminal-Bench 4.0 scores of 34% and 31.8% trail the model card's 42.3% (community reports)

More about DeepSeek