DeepSeek releases V4.1 Flash 1017 with open weights, stronger coding benchmarks and unchanged API pricing
DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, 2026, updating its V4.1 Flash model with improvements in reasoning, software engineering, autonomous agents and visual tasks. The company has deployed the checkpoint across its website, mobile application and API, while publishing model weights under the MIT license.
The update does not introduce a new architectural family or require developers to migrate to a different API identifier. Applications using `deepseek-flash` automatically access the new checkpoint, and DeepSeek says token pricing remains unchanged.
In DeepSeek's published evaluation, Terminal-Bench 4.0 rises from 31.2 to 42.3, ProgramBench from 20.3 to 29.1, and Automation-Bench from 54.8 to 60.5. The company also reports shorter outputs and faster completion of agentic tasks. Those results are developer-reported, so their operational significance depends on independent testing with matched tools and task budgets.
··········
DEEPSEEK-V4.1-FLASH-1017: WHAT CHANGED ON OCTOBER 17.
DeepSeek describes the October release as a minor checkpoint update with meaningful capability improvements. It builds on the V4.1 Flash family introduced in September, which uses an asymmetric Causal Encoder–Decoder architecture with sparse Mixture-of-Experts computation and visual understanding.
........
Specification | DeepSeek-V4.1-Flash-1017 |
Developer | DeepSeek |
Release | October 17, 2026 |
Previous checkpoint | DeepSeek-V4.1-Flash |
Architecture | Causal Encoder–Decoder MoE |
Previously disclosed MoE backbone | 552B parameters |
Active parameters during input | Approximately 8B |
Active parameters during output | Approximately 16B |
API identifier | deepseek-flash |
API context | 1 million tokens |
Maximum API output | 384K tokens |
Input | Text and images |
Thinking mode and tools | Supported |
Public weights and license | Available, MIT |
API pricing | Unchanged |
........
The disclosed MoE backbone parameter count differs from larger parameter totals in public model metadata, which can include other stored components. The October announcement emphasizes improved model behavior rather than a new architecture or a different input/output activation design.
··········
THE UPDATED MODEL DELIVERS DOUBLE-DIGIT GAINS ON TERMINAL-BENCH AND PROGRAMBENCH.
DeepSeek's direct comparison with the September checkpoint provides the clearest view of progress because both versions belong to the same family. All eleven evaluations listed by the developer improve, although the size of the gains varies considerably.
........
Benchmark | V4.1 Flash | Flash 1017 | Change |
HLE without tools | 36.8 | 42.6 | +5.8 |
Terminal-Bench 4.0 | 31.2 | 42.3 | +11.1 |
DeepSWE v1.1 | 74.2 | 77.0 | +2.8 |
ProgramBench | 20.3 | 29.1 | +8.8 |
NL2Repo-Bench | 65.4 | 68.2 | +2.8 |
Automation-Bench v1.0.6 | 54.8 | 60.5 | +5.7 |
Agents' Last Exam | 31.8 | 35.4 | +3.6 |
Apex Bench v1.1 | 64.6 | 72.1 | +7.5 |
Chartography with tools | 78.9 | 83.0 | +4.1 |
BabyVision with tools | 89.6 | 92.5 | +2.9 |
ZeroBench-main with tools | 49.0 | 51.0 | +2.0 |
........
Terminal-Bench gains 11.1 percentage points, equivalent to approximately 35.6% relative improvement over its previous score. ProgramBench's 8.8-point gain represents approximately 43.3% relative improvement. These calculations describe benchmark-score changes, not equivalent increases in overall intelligence or productivity.
The largest improvements concentrate on terminal execution, program construction and structured agent tasks, making this checkpoint particularly relevant to developers evaluating autonomous software workflows.
··········
CODING PERFORMANCE COMPARES FAVORABLY WITH GPT-5.6 SOL ON SELECTED TESTS.
DeepSeek also publishes comparisons with GPT-5.6 Sol, Claude Opus 5 and other models. The results are mixed rather than uniformly favorable to one system.
........
Benchmark | DeepSeek 1017 | GPT-5.6 Sol | Claude Opus 5 |
HLE without tools | 42.6 | 44.5 | 56.3 |
Terminal-Bench 4.0 | 42.3 | 39.9 | 51.8 |
DeepSWE v1.1 | 77.0 | 73.0 | 74.0 |
ProgramBench | 29.1 | 23.0 | 37.0 |
NL2Repo-Bench | 68.2 | 56.8 | 75.3 |
Automation-Bench | 60.5 | 45.8 | 50.3 |
Agents' Last Exam | 35.4 | 26.7 | 28.6 |
Apex Bench v1.1 | 72.1 | 54.3 | 65.8 |
........
Under the developer's reported conditions, DeepSeek exceeds GPT-5.6 Sol on seven of these eight benchmarks. Claude Opus 5 retains advantages on several reasoning and coding evaluations, including ProgramBench and Terminal-Bench.
DeepSeek specifies that Terminal-Bench uses the Claude Code harness, while DeepSWE uses Mini-SWE-V2. Harnesses can change task performance through tool access, retry policies and context handling, so these figures should not be interpreted as a fully independent or normalized ranking of commercial products.
··········
REINFORCEMENT LEARNING WITH SELF-FEEDBACK TARGETS COMPLETE PROJECT EXECUTION.
DeepSeek attributes part of the checkpoint's progress to reinforcement learning with self-feedback, especially for tasks requiring the construction of entire projects from broad instructions. Demonstrations include 3D modeling, web games and game-engine development, including a Portal 2-inspired Godot project and a representation of Beijing's Forbidden City.
Project-building agents face cumulative challenges: they must translate broad requirements into architecture, implementation, tests and revisions, while recognizing when an earlier choice has created downstream problems. Self-feedback can help identify and correct intermediate errors, but impressive demonstrations do not establish reliable autonomy across arbitrary repositories or production environments.
The useful measure is repeatable completion under a fixed time, tool and cost budget—not the visual sophistication of an isolated example.
··········
SHORTER OUTPUTS COULD REDUCE COSTS WITHOUT LOWER TOKEN PRICES.
DeepSeek says the October checkpoint improves the performance–cost trade-off by generating shorter responses and completing agentic tasks faster. Because the price per token remains unchanged, potential savings depend on reducing token consumption, retries or execution time.
Consider a hypothetical coding task using 100,000 uncached input tokens and 40,000 output tokens. If the updated model completes it with the same input and 25,000 output tokens, output usage falls 37.5%.
........
Illustrative component | Earlier workflow | Shorter workflow |
Input tokens | 100,000 | 100,000 |
Output tokens | 40,000 | 25,000 |
Input cost at $0.15/M | $0.015 | $0.015 |
Output cost at $0.60/M | $0.024 | $0.015 |
Total model cost | $0.039 | $0.030 |
Total reduction | — | 23.1% |
........
These figures are hypothetical rather than measured results for the new checkpoint. They also exclude search, code execution and retries. A shorter response only improves economics if it still completes the task correctly; an overly compressed answer that requires another attempt can cost more overall.
··········
DEEPSEEK-FLASH API PRICING REMAINS UNCHANGED.
DeepSeek maintains peak and off-peak rates, with discounts for eligible cached input. Prices below are in US dollars per million tokens.
........
API usage | Off-peak | Peak |
Input, cache hit | $0.003 | $0.006 |
Input, cache miss | $0.15 | $0.30 |
Output | $0.60 | $1.20 |
........
Peak periods are 01:00–04:00 UTC and 06:00–10:00 UTC Monday through Friday, excluding Chinese public holidays. Other periods, including weekends, use off-peak rates.
Off-peak rates halve per-token charges compared with peak periods. Eligible cached input at $0.003/M is 98% cheaper than uncached input at $0.15/M during off-peak hours, making cache utilization potentially significant for agents that repeatedly reuse long instructions or shared reference material. Actual savings depend on cache eligibility and prompt structure.
··········
THE CAUSAL ENCODER–DECODER ARCHITECTURE SEPARATES INPUT AND OUTPUT COMPUTATION.
V4.1 Flash uses an asymmetric Causal Encoder–Decoder design with sparse expert computation. The September architecture disclosure described a 552B MoE backbone activating approximately eight billion parameters while processing input and sixteen billion while generating output.
Prefill and decoding have different performance characteristics: input tokens can be processed together, whereas generated tokens depend sequentially on previous output. Optimizing these stages differently can reduce unnecessary computation and improve serving efficiency.
The family also emphasizes KV-cache compression. DeepSeek previously reported approximately one-quarter of the high-bandwidth memory and one-eighth of the SSD storage for its KV cache compared with the preceding generation. These are architectural-family claims from September, not newly introduced October improvements.
··········
THE OPEN-WEIGHT CHECKPOINT IS AVAILABLE UNDER THE MIT LICENSE.
DeepSeek has published V4.1-Flash-1017 through Hugging Face and ModelScope, with MIT licensing and reference components for inference and conversation encoding. The Hugging Face file listing is approximately 510 GB, while model metadata reports approximately 763B parameters across stored components.
The larger metadata total should not be confused with the previously described 552B MoE backbone. Stored model components, quantization and tensor representations affect the relationship between parameter counts and checkpoint size.
Open weights provide flexibility for customization and private deployment, but this is not a lightweight single-GPU model. Hosting requires substantial accelerator memory, compatible software and management of parallelism, attention states and concurrency. For many organizations, the hosted API may be more economical unless volume, privacy or customization requirements justify dedicated infrastructure.
··········
AGENT TOOLING AND ONE-MILLION-TOKEN CONTEXT REQUIRE PRACTICAL TESTING.
The repository includes reference components for encoding conversations, reasoning settings, tool calls and interleaved image content. DeepSeek also points to its deepseek-recipe toolkit, including Rust libraries with Python bindings. Correct message formatting matters because even a strong checkpoint can perform poorly when integrations mishandle tool boundaries or conversation history.
The hosted API advertises a one-million-token context window and a 384,000-token maximum output. These are API specifications rather than guarantees that every third-party deployment exposes the same limits or that retrieval remains reliable throughout a million-token input.
Large contexts can accommodate repositories and document archives, but developers should measure retrieval across distant positions, handling of contradictory sources and accuracy after repeated tool calls. Selective retrieval and context management may still be more economical than repeatedly sending complete document collections.
··········
SEARCH AND FETCH APIS ADD A NEW COST COMPONENT FOR AGENTS.
The October announcement also introduces public access to DeepSeek's Search and Fetch APIs, powered by the company's search infrastructure. Both endpoints are priced at $0.75 per 1,000 successful requests.
........
Service | Published price |
Search API | $0.75 per 1,000 calls |
Fetch API | $0.75 per 1,000 calls |
10,000 search calls | $7.50 |
10,000 fetch calls | $7.50 |
........
DeepSeek states that server-side Search and Fetch calls through supported agent integrations remain free for DeepSeek API users until billing begins on October 31, 2026, at 16:00 UTC, corresponding to November 1 in Beijing.
For research agents, search and retrieval requests can add to model token charges and tool execution costs. The correct operating metric is therefore total expenditure per completed workflow, incorporating requests, tokens, retries and elapsed execution time.
··········
WHAT THE UPDATE MEANS FOR ENTERPRISE CODING AND AUTOMATION.
The immediate benefit for existing users is access to a potentially stronger model without a new API identifier or higher per-token price. That convenience introduces a reproducibility concern: applications using a continuously updated alias can change behavior even when their own code remains identical.
For critical processes, organizations should run regression tests before broad deployment, comparing the old and new checkpoint on their actual repositories, documents and tool workflows. A coding assistant should be evaluated on passing tests and recovering from failures; a financial-analysis agent should be tested for arithmetic accuracy, reporting-period consistency and traceable source evidence.
........
Evaluation dimension | Recommended measurement |
Coding quality | Successful repository tasks |
Tool reliability | Valid calls and error recovery |
Efficiency | Time per completed workflow |
Token use | Input and output tokens per success |
Long context | Retrieval and evidence consistency |
Cost | Total expenditure per success |
Stability | Regression rate |
Security | Permissions and prompt-injection resistance |
........
Such testing is more informative than a general leaderboard because it measures the model under the organization's real constraints and operating environment.
··········
V4.1-FLASH-1017 IMPROVES AGENTIC CAPABILITIES WITHOUT A NEW PRICE TIER.
DeepSeek-V4.1-Flash-1017 is a checkpoint update rather than a new architectural generation. It retains the V4.1 Flash family's asymmetric sparse-computation design while delivering developer-reported gains in coding, terminal execution, automation and reasoning.
Terminal-Bench improves by 11.1 points, ProgramBench by 8.8 points and Apex Bench by 7.5 points, while DeepSeek reports shorter outputs and faster agentic task completion at unchanged token rates. MIT-licensed weights provide deployment flexibility, although the checkpoint remains expensive to host independently.
The commercial question is whether the updated model completes more demanding tasks with fewer resources at the same token price. That is a measurable advantage, and one that developers can now evaluate against their existing DeepSeek workloads.
··········
FOLLOW US FOR MORE.
DATA STUDIOS
datastudios.org
