Overview
DeepSeek-V4.1-Flash is the smallest model in DeepSeek's brand-new architecture series, released on September 10, 2026. Built on a novel Causal Encoder-Decoder architecture, it is designed for a higher capability ceiling, faster inference and greater throughput, with a path to scale to larger parameter counts.
The model is a 552B MoE that activates about 8B parameters for prefill and 16B for decoding, supporting 1M context with up to 384K output. Its most notable innovation is KV cache compression: the FP4-quantized global KV cache drops to roughly 890 bytes per token, about a quarter of the previous V4-Flash, dramatically lowering the memory and storage cost of long-context agent workloads.
It is also DeepSeek's first official model with native multimodal vision understanding, carrying forward the direction validated by the August Vision-Exp experimental release. The API model name is deepseek-flash, weights are open-sourced under MIT on Hugging Face, and third-party cloud platforms have onboarded it in parallel.
On 2026-10-17 DeepSeek shipped the DeepSeek-V4.1-Flash-1017 checkpoint, a weight-only update to the same checkpoint of about 763B total weights (552B backbone + 196B Engram memory, plus vision tower and embeddings) under MIT, with the deepseek-flash API name and prices unchanged; all 11 self-reported benchmarks improve, with DeepSWE v1.1 reaching 77.0 and per-task cost dropping from about $0.50 to $0.20, and Search and Fetch APIs also open. The 1017 build went on to take first among open models on Agent Arena on 2026-10-19, ranking 11th overall with a net gain of 5.1% across more than 7,500 real sessions at a median task cost of $0.20, and scored 45 on the Artificial Analysis Intelligence Index, six points above the prior version.
Key Features
- Native multimodal vision: Handles image input directly in a single model with no external vision module, scoring 78.9 on Chartography w/tools and 89.6 on BabyVision w/tools, turning the August experimental direction into a production model.
- Radically compressed KV cache: FP4 global KV cache at roughly 890 bytes per token, about a quarter of V4-Flash in HBM and an eighth in SSD storage, sharply lowering the hardware barrier for long-session agents and high-concurrency batch work.
- 1M context and 384K output: A 1M-token context window with up to 384K output, paired with minimal KV cache overhead, fits repository-scale understanding, very long documents and multi-turn agent tasks.
- Stronger agent capabilities: Terminal-Bench 2.1 90.6, DeepSWE v1.1 74.2, CyberGym 88.1, NL2Repo-Bench 65.4 and Automation-Bench 54.8, officially tested on the DeepSeek Harness framework at the max level. The 2026-10-17 checkpoint improves all 11 self-reported benchmarks, with DeepSWE v1.1 reaching 77.0, and the 1017 build took first among open models on Agent Arena on 2026-10-19, ranking 11th overall with a net gain of 5.1% across more than 7,500 real sessions at a median task cost of $0.20.
- MIT open source: Weights released under the MIT License on Hugging Face, free for commercial use and derivative work, continuing DeepSeek's fully open strategy. The 2026-10-17 checkpoint is a weight-only update of about 763B total weights (552B backbone + 196B Engram memory, plus vision tower and embeddings), also under MIT.
- Major API price cut: API prices were lowered in tandem, with cache-miss input from 1 yuan per million tokens (off-peak), plus peak/off-peak pricing and up to 2500 concurrent connections. The 2026-10-17 checkpoint keeps the deepseek-flash API name and prices unchanged ($0.15 input / $0.60 output per million tokens off-peak) and opens Search and Fetch APIs.
- Benchmark standing: Scored 45 on the Artificial Analysis Intelligence Index on 2026-10-19, six points above the prior version and tying GLM-5.3 for second among open-weight models, and leads the open-model board on Code Arena WebDev. On 2026-10-14 Artificial Analysis added it to AA-AgentPerf, where it leads on throughput in testing on GB300. ARC Prize published results on 2026-10-06: 72.9 percent on ARC-AGI-2 at $0.13 per task and 94.5 percent on ARC-AGI-1 at $0.07 per task.
Use Cases
- Long-context agent tasks: 1M context with a quarter of the KV cache overhead suits long-session agents, repository-scale analysis and bulk document processing
- Multimodal applications: chart understanding, visual QA and screenshot-to-action scenarios call the model directly without extra components
- Cost-sensitive scale-out: peak/off-peak pricing and a high concurrency limit make overnight off-peak batch jobs very cheap
- Self-hosted deployment: MIT license and low memory requirements let enterprises run higher concurrency on limited GPU resources
- Coding and R&D automation: Terminal-Bench, DeepSWE and CyberGym results show strength in terminal operations, software engineering and security tasks, with Fireworks AI reporting DeepSWE results on par with GPT-6 Astra at roughly one fifteenth of the per-run cost on 2026-09-14
Pros
- Official and third-party testing agree it outperforms DeepSeek's own flagship V4 Pro on performance, cost, speed and total time
- Industry-leading KV cache compression cuts long-context agent memory cost by an order of magnitude
- Native vision understanding unifies multimodal and text tasks in one model
- Fully open under MIT with no commercial restrictions
- Extremely low API pricing, with cache-miss input at 1 yuan per million tokens off-peak
- Available day-one across DeepSeek API, SiliconFlow and other channels
- The 2026-10-17 checkpoint took first among open models on Agent Arena on 2026-10-19 and scored 45 on the Artificial Analysis Intelligence Index
- Independent evaluation from Fireworks AI on 2026-09-14 reported DeepSWE results on par with GPT-6 Astra at roughly one fifteenth of the per-run cost
Pricing
Peak/off-peak API pricing (per million tokens): cache-hit input 0.02 yuan off-peak and 0.04 yuan peak; cache-miss input 1 yuan off-peak and 2 yuan peak; output 4 yuan off-peak and 8 yuan peak. Peak hours are weekdays 9:00-12:00 and 14:00-18:00 Beijing time. Concurrency limit 2500. Weights are MIT-licensed with no fee for self-hosting. The 2026-10-17 checkpoint keeps the deepseek-flash API name and prices unchanged ($0.15 input / $0.60 output per million tokens off-peak), while per-task cost fell from about $0.50 to $0.20.
Summary
DeepSeek-V4.1-Flash is the most important open-source model update from China in 2026. Its new architecture delivers higher capability and a dramatic cost reduction at the same time, and DeepSeek is confident enough to retire its own flagship V4 Pro in its favor. The 2026-10-17 checkpoint pushed it further: it took first among open models on Agent Arena on 2026-10-19 and scored 45 on the Artificial Analysis Intelligence Index, while independent evaluations from Fireworks AI and Agent Arena place its coding-agent results alongside far costlier frontier models. For teams needing long-context agents, native vision and low-cost scale-out, it is a top pick in the open-source field; for API users, peak/off-peak pricing plus cache discounts push costs to the lowest tier in the industry.
Version History
- 1017 build takes first among open models on Agent Arena and scores 45 on the AA Intelligence Index (2026-10-19): Arena reported on October 19 that DeepSeek-V4.1-Flash-20261017 (Max) became the top open model on Agent Arena, ranking 11th overall with a net gain of 5.1% across more than 7,500 real sessions. Its median task cost is $0.20, which the team compares to roughly 37% of GPT-6.1 Sol's improvement at 27% of the cost, with Confirmed Success up 11.5%. On the Artificial Analysis Intelligence Index it scored 45, six points above the prior version, tying GLM-5.3 for second among open-weight models behind MiMo-V2.6-Pro, and it leads the open-model board on Code Arena WebDev.
- DeepSeek-V4.1-Flash-1017 checkpoint released (2026-10-17): Ships DeepSeek-V4.1-Flash-1017, a weight-only update to the same checkpoint of about 763B total weights (552B backbone + 196B Engram memory, plus vision tower and embeddings) under MIT; the deepseek-flash API name and prices are unchanged ($0.15 input / $0.60 output per million tokens off-peak); all 11 self-reported benchmarks improve, with DeepSWE v1.1 reaching 77.0 (above Claude Opus 5 at 74.0) and per-task cost dropping from about $0.50 to $0.20; also opens Search and Fetch APIs.
- Added to the Artificial Analysis AA-AgentPerf benchmark, leading on GB300 throughput (2026-10-14): On October 14, 2026, Artificial Analysis added DeepSeek V4.1 Flash to AA-AgentPerf, its agent performance benchmark, where the model leads on throughput in testing on GB300. AA-AgentPerf looks at how efficiently agents execute real tasks, with throughput and latency among its key measurements.
- ARC-AGI results published (2026-10-06): ARC Prize published DeepSeek V4.1 Flash results on ARC-AGI (Verified): 72.9 percent on ARC-AGI-2 at $0.13 per task and 94.5 percent on ARC-AGI-1 at $0.07 per task. That is 11.5 and 5.5 points above the best V4 Flash scores, at roughly 250 percent higher cost per task
- Fireworks benchmark: DeepSWE results on par with GPT-6 Astra at about one fifteenth of the cost (2026-09-14): Fireworks AI published an independent evaluation of DeepSeek-V4.1-Flash, reporting DeepSWE results on par with GPT-6 Astra at roughly one fifteenth of the per-run cost. Fireworks also lists the model among its recommended inference targets, emphasizing its cost advantage on coding-agent workloads.
- Ranks 3rd among open models on Agent Arena at a median cost of about $0.07 per task (2026-09-14): The Agent Arena leaderboard places DeepSeek-V4.1-Flash (Max) third among open models, with a median cost of about $0.07 per task. The model also scores 40 on Artificial Analysis Intelligence Index v4.3, putting it in the leading open-weight group alongside GLM-5.3 Flash and Qwen3.8 2.4T, with output speed of 214.4 tokens/s among the highest in its class.
- DeepSeek-V4.1-Flash officially released, open-sourced and live on the API (2026-09-10): First release of the smallest model in the new architecture family: 552B MoE (8B prefill / 16B decode active), native vision, 1M context and FP4 KV cache at ~890 bytes per token. API model name deepseek-flash; legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp names are retired and routed to the new model, and V4 Pro will be phased out (requests routed to V4.1 Flash after 12:00, Sep 14). Open-sourced under MIT with API prices cut in tandem.