DeepSeek’s V4.1 Flash update jumps 11 points on Terminal-Bench 4.0, MIT-licensed

DeepSeek has published DeepSeek-V4.1-Flash-1017 on Hugging Face, an update to its V4.1-Flash model. The biggest listed jump is on Terminal-Bench 4.0, where it scores 42.3, up from 31.2 for the previous version.

What is new in V4.1-Flash-1017

The model card calls the release a minor update to DeepSeek-V4.1-Flash. According to DeepSeek, it improves reasoning, coding, agentic tasks and visual benchmarks, and it reduces task execution time and output length, which the company ties to a better balance of performance and cost across agent harnesses. The card gives no exact figures for those time and length savings.

The listing shows a model of 763B parameters in mixed precision (BF16, F32, F8_E4M3 and I8), tagged for image-text-to-text use. DeepSeek recommends a 1M-token context window and a maximum output of at least 256K tokens, with temperature and top_p both at 1.0. The repository and weights are under the MIT License. The card links a technical report and cites it under the title “Pushing the Limits of KV Cache Compression.”

Usage instructions cover Transformers, vLLM, SGLang and Docker Model Runner, and one quantized version is listed for llama.cpp, LM Studio, Jan and Ollama. There is no Jinja chat template; instead, a Python reference implementation handles multi-turn chat, tool calling, thinking mode, numeric reasoning effort and interleaved images. No inference provider hosts the model yet, according to the page.

The benchmark numbers

DeepSeek compares the new build with the previous Flash release and with K3, GLM-5.3, GPT-5.6 Sol and Opus 5. All figures below come from the model card.

• Terminal-Bench 4.0 (Claude Code harness): 42.3, against 31.2 before, 39.9 for GPT-5.6 Sol and 51.8 for Opus 5.
• DeepSWE v1.1: 77.0, ahead of every listed model, including GPT-5.6 Sol (73.0) and Opus 5 (74.0).
• Apex Bench v1.1: 72.1, against 65.8 for Opus 5.
• Automation-Bench v1.0.6: 60.5 and Agents’ Last Exam: 35.4, both first in the table.
• HLE without tools: 42.6, up from 36.8, but behind Opus 5 at 56.3.
• Vision with tools: Chartography 83.0, BabyVision 92.5, ZeroBench-main 51.0; Opus 5 is slightly ahead on the first two and GPT-5.6 Sol on the third.

The pattern is clear: the update leads on several agentic and software-engineering tests but still trails Opus 5 on HLE, Terminal-Bench and ProgramBench (29.1 against 37.0). These are DeepSeek’s own reported results, and the card does not say they were independently reproduced.

Why it matters

The technical angle is the cache. When a model reads a 1M-token prompt, the memory it keeps for every token (the KV cache) becomes the main cost of serving it, and that is why DeepSeek’s report title focuses on compressing it. Our earlier coverage of the April V4 launch described the family as a Pro model for heavy tasks and a lighter Flash model aimed at lower cost, with a one-million-token context and lower compute and memory demands. V4.1-Flash-1017 keeps that design and now pushes agent scores up, which is where long context pays off: coding agents that carry whole repositories and long tool histories.

Shorter outputs matter for the same reason. Agent runs repeat model calls many times, so fewer tokens per step cuts both latency and bill, and that is the trade-off DeepSeek highlights rather than raw scores alone.

The release also lands in a politically charged setting. In April, a White House memo from science adviser Michael Kratsios accused Chinese players of large-scale distillation of American models, as we reported in our report on the planned crackdown. The same report noted that industry insiders suggest DeepSeek’s low-cost model leaned on distilling ChatGPT or Claude, which it presented as a suggestion rather than a confirmed finding. A freely licensed 763B model that sits close to US frontier systems on several agent tests keeps that argument alive, whatever its origin.

With an MIT license, companies can download, modify and self-host the weights without a separate agreement. Real-world performance will only be clear once independent testers and hosting providers run the model; for now the numbers are the lab’s own.