DeepSeek keeps shipping. One week after the September V4.1-Flash release, the company is back with DeepSeek-V4.1-Flash-1017, a point update that claims gains on all eleven benchmarks in its comparison table. The weights are open under the MIT license, the API price has not moved, and existing API users get the new model automatically. It is the kind of release that is easy to scroll past and slightly annoying to have ignored, if you run agent workloads on a budget.
What changed
The model card lists 763B parameters, multimodal input (image and text), a 1M token context window through the API, and up to 384K output tokens. The technical report behind the V4.1-Flash line is titled “Pushing the Limits of KV Cache Compression,” which tells you where DeepSeek thinks the leverage is: not raw scale, but making long-context agentic work cheap enough to run in loops.
The headline benchmark jumps are Terminal-Bench 4.0 going from 31.2 to 42.3 and ProgramBench from 20.3 to 29.1. Agentic scores also moved: DeepSWE v1.1 at 77.0, Automation-Bench v1.0.6 at 60.5, Agents’ Last Exam at 35.4, Apex Bench v1.1 at 72.1. DeepSeek trained the update with reinforcement learning using self-feedback to improve end-to-end project building, and the demos include a 3D model of the Forbidden City and a Portal 2 recreation in Godot.
Two caveats before you quote those numbers at anyone. First, they are all vendor-reported; nobody independent has rerun them yet. Second, the comparison is not strictly like-for-like: the Terminal-Bench 4.0 runs used the Claude Code harness while the previous card used DeepSeek Harness, so part of the “gain” could be harness differences. I would treat the direction as real and the magnitude as soft.
Vision is in the box too. The API serves the model with image input and vision-exp capabilities under the same endpoint, and the legacy model names deepseek-v4-flash and deepseek-v4-flash-vision-exp now route to this snapshot at Flash prices. If you were running the older model, you are already on it.
The honest competitive picture
On DeepSeek’s own table, Claude Opus 5 still leads six of the eleven rows, including HLE at 56.3 and Terminal-Bench 4.0 at 51.8. DeepSeek leads the rest. So this is not a “we beat the frontier” release. It is a “we beat you on cost per completed task” release, and given what the API costs, that is a serious pitch.
The API price is unchanged under the deepseek-flash name: off-peak, $0.15 per 1M cache-miss input tokens, $0.003 per 1M cache-hit input tokens, and $0.60 per 1M output tokens. Peak hours (01:00-04:00 and 06:00-10:00 UTC, Monday to Friday) double those rates. DeepSeek says the new model produces shorter outputs and finishes tasks faster, a performance-cost Pareto improvement, though it published no specific speed figures. Shorter outputs matter more than people realize: for an agent that burns thousands of turns, output tokens are usually the biggest line on the bill.
The bigger move: Search and Fetch APIs
Bundled with the model release, DeepSeek opened its Search and Fetch APIs to all users at $0.75 per 1,000 successful requests each. Until billing starts on October 31 at 16:00 UTC, calls through DeepSeek Harness, Claude Code, and Codex are free. Read that as DeepSeek moving up the stack. Model, harness, search, fetch: that is an agent platform, not just a model provider, and the Claude Code and Codex integrations show exactly which developer base it is courting.
If you want to run it yourself
Self-hosting has one quirk worth planning around. There is no Jinja chat template in the release. Instead, DeepSeek ships a Python reference encoder (encoding.py) plus deepseek-recipe, a set of Rust libraries with Python bindings for prompt encoding and response parsing. If your serving stack assumes a chat_template.jinja, budget an afternoon for the integration. Community quantizations are already appearing on Hugging Face, and third-party provider DeepInfra already lists the model at $0.20 per 1M input and $0.60 per 1M output.
The model card recommends temperature 1.0, top_p 1.0, and max_tokens of at least 256K. DeepSeek also recommends DeepSeek Harness v0.2.1-rc.1 or later, and says the model is optimized for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy.
What I would actually do
- Running agents on the API: schedule batch runs off-peak to halve token costs, and lean on the $0.003 cache-hit rate for repeated long contexts. The upgrade is already live under deepseek-flash, so pin the model version if your evals matter.
- Curious about the Search and Fetch APIs: test them before October 31 while they are free, then decide if the pricing works for your tooling.
- Self-hosting: start from the encoding.py reference or deepseek-recipe toolkit, and do not expect drop-in Jinja compatibility.
The pattern across DeepSeek’s last few releases is consistent: open weights at the frontier-adjacent tier, aggressive pricing aimed at agent loop economics, and now tooling APIs to keep developers inside its stack. Whether the benchmarks hold up under independent testing matters less than the direction of travel. Cheap, capable, open agent models with 1M context change what you can afford to automate, and the competition has to answer that.