DeepSeek V4.1 Flash 1017 is the October 17, 2026 snapshot of DeepSeek V4.1 Flash — a 552B-parameter sparse Mixture-of-Experts model built on DeepSeek's Causal Encoder-Decoder (CED) architecture, activating only ~8B parameters during prefill and ~16B during decode. It pairs a 1,048,576-token context window with native image input, controllable reasoning effort, MIT-licensed open weights, and an extremely low price ($0.15/1M in, $0.60/1M out), positioning it as a cost-dominant agentic/coding workhorse.
Ultra-Fast (~95-130 tps single-stream decode at 128K-512K context; ~6,000 tps prefill and ~17s TTFT at ~105K tokens on 8x-accelerator local deployments)
Verified Frontier Benchmarks
Empirical Evaluation Results
Evaluated across standard and verified harnesses
Terminal-Bench 4.0
coding
42.3%
Terminal-Bench 3.0
coding
43.3%
DeepSWE v1.1
coding
77.0%
Automation-Bench v1.0.6
agentic
60.5
Agents' Last Exam
agentic
35.4
Apex Bench v1.1
reasoning
72.1
ProgramBench
coding
29.1
NL2Repo-Bench
coding
68.2
Architectural Profile & Capabilities
DeepSeek V4.1 Flash 1017 (deepseek/deepseek-v4.1-flash-1017) is a dated snapshot served by DeepSeek under the `deepseek-flash` API model name. It supersedes the earlier V4.1 Flash release (September 2026) and, per DeepSeek's own changelog and companion model cards, delivers broad coding and agentic gains — notably lifting Terminal-Bench 4.0 from 31.2 to 42.3 and DeepSWE v1.1 from 74.2 to 77.0 at the same price point.
ARCHITECTURE: The model is a sparse Mixture-of-Experts (MoE) system with a reported 552B backbone parameters. Uniquely, it is DeepSeek's first model built on a Causal Encoder-Decoder (CED) architecture — a 40-layer Transformer split into a 20-layer causal encoder and a 20-layer decoder. The asymmetric activation pattern activates roughly 8B parameters during input/prefill and 16B during output/decode, which is why the launch coverage emphasizes cheaper prompt reading versus writing. The specific attention kernel (MLA/GQA/DSA) is not established in the available documentation and should not be assumed. The model supports continuously controllable reasoning/thinking effort and generates text autoregressively while accepting images.
CONTEXT & SERVING: The 1,048,576-token (1M) context window is documented and is used for the coding-agent evaluations (reported with the DeepSeek Harness Minimal mode and Claude Code harness for Terminal-Bench 4.0). Output is capped at up to 384,000 tokens. A third-party local deployment on eight CMP 170HX accelerators measured ~117 tokens/s single-stream decode at 128K context, ~96 tokens/s at 512K, ~560 tokens/s aggregate across eight concurrent streams at 128K, and ~6,066 tokens/s prefill around 105K tokens with ~17s time-to-first-token. These are hardware- and implementation-specific figures, not provider SLA numbers.
POSITIONING: Compared to the October 2026 frontier (GPT-6 Astra, Claude Opus 5, Gemini 3.8 Pro, DeepSeek V4 Pro, Kimi K3), V4.1 Flash 1017 is not a top-end reasoning model — third-party coverage consistently notes Claude Opus 5 retains advantages on HLE and the hardest reasoning/terminal tasks. Its value proposition is elite agentic/coding throughput-per-dollar (roughly 30-40x cheaper than GPT-5.6 Sol-class pricing), open MIT weights, and a 1M context. Benchmark coverage remains concentrated on DeepSeek's own agent benchmark suite; most standardized third-party indices (AA Intelligence Index, SWE-bench Verified/Pro, GPQA Diamond, HLE, AIME 2026, BrowseComp, ARC-AGI-2, MRCR v2) were NOT verified for the exact 1017 snapshot and are therefore omitted.
Architecture Type: Sparse Mixture-of-Experts (MoE) on Causal Encoder-Decoder (CED) Transformer
552B backbone parameters; asymmetric activation of ~8B parameters during prefill/input and ~16B during decode/output. 40-layer Transformer split into a 20-layer causal encoder and 20-layer decoder (DeepSeek's first CED-architecture model). Supports controllable reasoning/thinking effort. Accepts text and image input and generates text autoregressively. 1,048,576-token context window with up to 384,000-token output. Exact attention kernel (MLA/GQA/DSA) not verified. MIT-licensed open weights released alongside the API.
Terminal and shell automation / computer-use agents (Terminal-Bench 4.0, Automation-Bench)
Long-context document, codebase, and log analysis leveraging the 1M-token window
High-volume content processing, classification, and extraction at $0.15/1M input pricing
Multimodal (image + text) understanding tasks such as screenshot/UI and diagram interpretation
Open-weight self-hosting and fine-tuning under the MIT license for sovereign/on-prem deployments
Operational Caveats & Boundaries
•Most standardized third-party benchmarks (AA Intelligence Index, SWE-bench Verified/Pro, GPQA Diamond, HLE with/without tools, AIME 2026, BrowseComp, ARC-AGI-2, MRCR v2, FrontierMath) are NOT verified for the exact 1017 snapshot — published scores cluster around DeepSeek's own agent suite, so cross-vendor comparison is limited.
•Prefill is 2x the active-parameter cost of decode (8B active prefill vs 16B active decode); very long prompts are far more expensive and slower than equivalent generation, and TTFT grows sharply at ~100K+ token inputs (~17s observed locally).
•Throughput and latency figures currently come from third-party local hardware deployments (8x CMP 170HX), not from DeepSeek/OpenRouter SLA telemetry — production API TPS and percentile latency are unverified.
•Some benchmark numbers (e.g. Terminal-Bench 4.0) depend heavily on the harness used (Claude Code frame for TB4.0, DeepSeek Harness Minimal mode with 1M context); scores are not directly comparable across different agent scaffolds.
•As a Flash-tier model, it trails flagship reasoning models (e.g. Claude Opus 5) on the hardest unaided knowledge/reasoning and terminal tasks — third-party reviews explicitly note it 'loses on unaided knowledge and reasoning.'
•The output cap (~384K tokens) is distinct from the 1M context ceiling and can be reached before the context limit in long-generation agent loops.
ARMES Zero Data Retention
Calls to v4.1 Flash 1017 are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by DeepSeek.
