DeepSeek published DeepSeek-V4.1-Flash-1017 on Hugging Face on 17 October. The model card calls it "a minor update to DeepSeek-V4.1-Flash", with gains in reasoning, coding, agentic tasks and visual benchmarks. It also claims shorter task execution time and shorter output, which would mean lower cost per task.
The benchmark deltas are not minor. If you route agent traffic to a Flash-tier model, this is worth a re-evaluation. I covered the earlier V4 Flash 0731 release when it landed. The 1017 card compares only against V4.1-Flash, not 0731.
The benchmark table, side by side
All numbers below come from DeepSeek's model card, so treat them as vendor-reported.
| Benchmark | V4.1-Flash-1017 | V4.1-Flash | K3 | GLM-5.3 | GPT-5.6 Sol | Opus 5 |
|---|---|---|---|---|---|---|
| HLE w/o tools | 42.6 (44.0†) | 36.8 (39.1†) | 43.5 | 42.0† | 44.5 | 56.3 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | 12.6 | 37.9 | 39.9 | 51.8 |
| DeepSWE v1.1 | 77.0 | 74.2 | 67.5 | 66.9 | 73.0 | 74.0 |
| ProgramBench | 29.1 | 20.3 | 17.5 | 19.0 | 23.0 | 37.0 |
| NL2Repo-Bench | 68.2 | 65.4 | 58.0 | 58.0 | 56.8 | 75.3 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | 46.7 | 48.8 | 45.8 | 50.3 |
| Agents' Last Exam | 35.4 | 31.8 | 27.6 | 28.5 | 26.7 | 28.6 |
| Apex Bench v1.1 | 72.1 | 64.6 | 47.8 | 56.6 | 54.3 | 65.8 |
| Chartography (w/tools) | 83.0 | 78.9 | 68.1 | not stated | 79.9 | 84.0 |
| BabyVision (w/tools) | 92.5 | 89.6 | 85.7 | not stated | 88.9 | 94.1 |
| ZeroBench-main (w/tools) | 51.0 | 49.0 | 41.0 | not stated | 53.0 | 52.0 |
† marks the text-only subset of HLE.
The biggest jumps over the previous V4.1-Flash are Terminal-Bench 4.0 (+11.1), ProgramBench (+8.8) and Apex Bench (+7.5). The smallest are DeepSWE and NL2Repo at +2.8 each.
Where it beats Opus 5 and where it does not
On the card's numbers, Flash-1017 is ahead of Opus 5 on four agentic benchmarks: DeepSWE v1.1 (77.0 vs 74.0), Automation-Bench (60.5 vs 50.3), Agents' Last Exam (35.4 vs 28.6) and Apex Bench (72.1 vs 65.8).
It trails on the rest. Terminal-Bench 4.0 is 42.3 against 51.8, ProgramBench 29.1 against 37.0, NL2Repo 68.2 against 75.3, and HLE 42.6 against 56.3. Opus 5 also leads on Chartography and BabyVision, and on ZeroBench GPT-5.6 Sol is ahead at 53.0.
The harness matters. The card says Terminal-Bench 4.0 was run with the Claude Code harness and DeepSWE v1.1 with the Mini-SWE-V2 harness. Your own harness may land elsewhere. That is the argument for running your own tasks before switching.
Serving it: weights, license, sampling
The weights are MIT-licensed. Hugging Face lists 763B total parameters across BF16, F32, F8_E4M3 and I8 tensors, with a total file size of 510,304,178,606 bytes. That is roughly 510 GB, so this is a multi-GPU or big-memory-box deployment, not a single-card one. Hugging Face lists 31 quantizations and 5 finetunes in the model tree already.
The card's quick-start commands:
pip install vllm
vllm serve "deepseek-ai/DeepSeek-V4.1-Flash-1017"
pip install sglang
python3 -m sglang.launch_server \
--model-path "deepseek-ai/DeepSeek-V4.1-Flash-1017" \
--host 0.0.0.0 --port 30000
Recommended sampling parameters:
| Parameter | Value |
|---|---|
| temperature | 1.0 |
| top_p | 1.0 |
| context_window | 1M tokens |
| max_tokens | 256K or more |
The max_tokens guidance is the one to check in your gateway. Many proxies cap output far below 256K, and that would silently truncate long agent turns.
The prompt-format catch
This release ships no Jinja chat template. That breaks the usual assumption that apply_chat_template just works. Instead, the encoding folder has a self-contained Python reference, encoding.py, with test cases for multi-turn conversations, tool calling, thinking mode, numeric reasoning effort, mid-conversation system messages and interleaved image content.
For production, DeepSeek also released deepseek-recipe, a set of Rust libraries with Python bindings. It converts Messages, Chat Completions and Responses API requests into a Conversation format and encodes them into V4 and V4.1 prompts or token IDs. It also parses model output back into complete or streamed responses, covering thinking, tool calls and images. Inference, tool execution and HTTP transport are left to you.
If you run your own server, I would test tool calling and thinking mode against that reference encoder first. Template drift is a classic source of quiet quality loss.
What the card does not tell you
- Pricing and API model IDs: not stated on the model card. Check DeepSeek's API docs before assuming a drop-in swap.
- Output-length and latency savings: claimed, but no numbers are given.
- Hosted access: Hugging Face says no Inference Provider deploys it yet.
- Comparison to the 0731 checkpoint: absent. The baseline is V4.1-Flash.
My plan is to replay a sample of real agent traces against Flash-1017 and compare tokens per completed task, not just pass rate. That is where the "lower cost" claim will either hold or not.
Sources: