DeepSeek Refreshed V4.1-Flash Eight Days Later: 763B Parameters, MIT License, and a KV-Cache Squeeze

Eight days after V4.1-Flash first landed with our verdict of “good enough frontier parity,” DeepSeek has shipped a refreshed checkpoint, and the model card for DeepSeek-V4.1-Flash-1017, published October 17, reads less like a patch and more like a progress report. The headline internal gains are concentrated exactly where agent developers feel them: Terminal-Bench 4.0 jumped from 31.2 to 42.3, ProgramBench from 20.3 to 29.1, and DeepSWE v1.1 rose to 77.0, the highest number listed on the card against every named competitor, including Opus 5 and GPT-5.6 Sol. Bloomberg’s summary of the release: stronger benchmark performance, faster task completion, and shorter outputs.

What the card actually claims

The architecture is unchanged in shape: 763 billion total parameters in a deepseek_v41 multimodal (image-text-to-text) design, sharded across BF16, FP8, and INT8 tensor types into roughly 510 GB of weights, with a 1M-token context window and a recommendation to run max_tokens at 256K or above. The citation attached to the repo names the research focus: “Pushing the Limits of KV Cache Compression,” which explains where the speed and cost-per-task improvements come from; for long agentic rollouts, cache footprint is the binding constraint, and compressing it is worth more than another point on a static benchmark. Sampling guidance is notably hands-off: temperature 1.0, top_p 1.0.

Against the closed frontier, the gaps persist but narrow. HLE without tools sits at 42.6 versus Kimi K3’s 43.5, GPT-5.6 Sol’s 44.5, and Opus 5’s 56.3; Opus 5 also still leads Terminal-Bench at 51.8 versus 42.3 (different Terminal Bench versions are in circulation, and on the v2.1 scoring used in Reflection’s Beam card this week, DeepSeek V4.1 Flash still held the top open-weights mark at 90.6). Either way, DeepSeek remains the open-weights agent model to displace.

The distribution story is the underrated part

Two release details matter as much as the benchmarks. First, the license: MIT, for weights and code, with no commercial-use strings. That puts V4.1-Flash-1017 one step more permissive than Meta’s Muse family, whose custom license carries limited commercial terms, and it matters for anyone embedding the model in a product. Second, ergonomics: there is no Jinja chat template in the repo. Prompt encoding for multi-turn conversations, tool calls, thinking mode, and images runs through a Python reference implementation (encoding.py) or a Rust toolkit with Python bindings (deepseek-recipe), an unusual choice that trades drop-in compatibility for exactness. The ecosystem has already absorbed the friction: 32 community quantizations exist for llama.cpp, Ollama, and LM Studio, though no inference provider serves the model on Hugging Face directly.

Why a same-week refresh is the signal

Shipping a checkpoint refresh eight days after the base release breaks the industry’s seasonal rhythm, and it is the second time this month a lab has treated agents as the fast-moving target: Beam’s team and DeepSeek’s are both publishing agent-harness numbers (DeepSeek’s card explicitly claims movement on “the performance-cost Pareto frontier across multiple agent harnesses”) rather than chat benchmarks. The practical read for engineering teams: if you standardized on V4.1-Flash last week, re-run your evals before re-pricing anything, because the agent numbers moved 10 to 35% in eight days. And watch the KV-cache work specifically; if DeepSeek publishes the compression technique behind the shorter outputs, it lands at exactly the moment distillation and efficiency disputes are already reshaping how open labs compete with closed ones.