DeepSeek-V4.1-Flash-1017: Flash Model Matches Flagships on Agent Benchmarks, Cuts KV Cache 4x

DeepSeek launched DeepSeek-V4.1-Flash-1017 over the weekend, a flash-tier model that delivers flagship-level performance on agent-centric benchmarks while significantly lowering inference cost.

Benchmark Results

The model card lists 11 benchmarks; the 1017 version improves on all of them versus the original V4.1-Flash:

DeepSWE v1.1 : 77.0 — highest among all listed models, ahead of GPT-5.6 Sol (73.0) and Opus 5 (74.0).

Automation-Bench : 60.5 (1st place).

Agents' Last Exam : 35.4 (1st place).

Apex Bench : 72.1 (1st place).

On HLE (42.6), Terminal-Bench 4.0 , ProgramBench , NL2Repo-Bench , and several visual tasks, Opus 5 still leads.

Architecture: Causal Encoder-Decoder with Sparse Activation

The backbone uses a 40-layer Transformer split into a 20-layer encoder + 20-layer decoder (Causal Encoder-Decoder). Total backbone parameters: 552B . During prefill only 8B parameters are activated per token; during decode 16B are activated. This sparse activation makes the model especially friendly for "input-heavy" agent scenarios.

KV Cache Compression: Pushing the Limits

The model's subtitle, Pushing the Limits of KV Cache Compression , reflects three key techniques:

CSA2 sparse attention + FP4 KV cache : global KV cache compressed to 890 bytes/token , roughly 1/4 of the original V4-Flash size.

SWA Bounded Replay : persistent KV cache further reduced to about 1/8 of V4-Flash.

Multimodal & Long Context

Native multimodal support.

1M token context window.

Trained from scratch on 45T tokens of multimodal corpus.

Adjustable Reasoning Intensity

Reasoning strength can be tuned continuously from 1 to 100 — lower for cost savings, higher for accuracy.

Open-Source Release Details

No Jinja chat template; instead a self-contained Python reference implementation ( encoding.py) covering multi-turn chat, tool calling, thinking mode, numeric reasoning intensity, mid-conversation system messages, and interleaved image-text.

For production, deepseek-recipe (Rust library + Python bindings) unifies Messages, Chat Completions, and Responses into V4/V4.1 prompts.

Recommended Sampling Parameters

temperature 1.0

top_p 1.0

Context window: 1M

max_tokens ≥ 256K

Search & Fetch API: Real-Time Web for Agents & RAG

DeepSeek API now includes a Search & Fetch suite:

Search powered by DeepSeek's proprietary search engine, returns structured results.

Fetch scrapes a single page and converts it to Markdown.

Workflow supports operators site:, lang:, after:; up to 50 results per request.

Pricing: $0.75 per 1,000 successful requests ; empty results and failed requests are free.

Server-side tool calling via DeepSeek Harness , Claude Code , Codex is free until 2026-11-01 00:00 Beijing time .

In summary, a flash-class model now competes head-to-head with flagship models on multiple agent benchmarks while cutting KV cache costs dramatically, and the bundled Search & Fetch API equips agents with both "brain" and "eyes" for real-time information retrieval.