DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, 2026 as open weights under the MIT License. DeepSeek's own model card calls it "a minor update" to V4.1 Flash, then shows a benchmark table where Terminal-Bench 4.0 moves from 31.2 to 42.3 and Apex Bench v1.1 from 64.6 to 72.1. That is a large jump for a point release, and every number in the table is self-reported.
The model is already live on DeepSeek's website, app and API, according to the official announcement thread on X. This post covers what changed, what it costs, how it compares in DeepSeek's own table, and what to verify before you switch production traffic.
TL;DR: the questions people ask first
| Question | Answer |
|---|---|
| What is it? | An update to DeepSeek V4.1 Flash, released October 17, 2026 as open weights |
| License? | MIT, for the repository and weights (per the model card) |
| Size? | About 763B parameters listed in the Hugging Face metadata, mixed FP8 and int8, roughly 510 GB of files |
| Context window? | 1M tokens; recommended max output of at least 256K tokens, API maximum 384K |
| Vision? | Yes. The API docs list vision as supported for deepseek-flash and not for deepseek-v4-pro |
| API price? | $0.15 input (cache miss) and $0.60 output per million tokens off-peak; double at peak |
| Where do the benchmarks come from? | DeepSeek's own table. No independent reproduction found at release |
| What else shipped? | Search and Fetch APIs open to everyone at $0.75 per 1,000 calls each |
What changed from the previous Flash
DeepSeek's description is short: stronger benchmark performance, faster task completion and shorter outputs than the previous version, which it calls "a Pareto improvement in performance and cost." The company says it used reinforcement learning with self-feedback to improve end-to-end project development, from game creation to 3D modeling, and shows GIF demos in the X thread.
The announcement also says the model is optimized for DeepSeek Harness, Tencent CodeBuddy, WorkBuddy and other agent harnesses, and recommends upgrading to DeepSeek Harness v0.2.1-rc.1 or later. That matters: agent benchmarks depend heavily on the harness, and DeepSeek states that it used the Claude Code harness for Terminal-Bench 4.0 and the Mini-SWE-V2 harness for DeepSWE v1.1.
For background on the model line, see our earlier coverage of the V4.1 Flash API beta and its new multimodal architecture and the roughly 75 percent KV cache memory reduction that makes a 1M-token context affordable to serve.
The benchmark table, with deltas
Row of passing test dots showing DeepSeek V4.1 Flash 1017 benchmark gains over the previous Flash
The table below is DeepSeek's, copied from the model card, with the change from the previous Flash computed by us.
| Benchmark | V4.1 Flash 1017 | V4.1 Flash | Change |
|---|---|---|---|
| HLE without tools | 42.6 (44.0 text-only subset) | 36.8 (39.1) | +5.8 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | +11.1 |
| DeepSWE v1.1 | 77.0 | 74.2 | +2.8 |
| ProgramBench | 29.1 | 20.3 | +8.8 |
| NL2Repo-Bench | 68.2 | 65.4 | +2.8 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | +5.7 |
| Agents' Last Exam | 35.4 | 31.8 | +3.6 |
| Apex Bench v1.1 | 72.1 | 64.6 | +7.5 |
| Chartography (with tools) | 83.0 | 78.9 | +4.1 |
| BabyVision (with tools) | 92.5 | 89.6 | +2.9 |
| ZeroBench-main (with tools) | 51.0 | 49.0 | +2.0 |
The biggest gains are on terminal and agent tasks (Terminal-Bench, ProgramBench, Apex), the area DeepSeek says it trained for with self-feedback reinforcement learning. The smallest are on DeepSWE and vision, where the previous model was already strong.
How it compares with other models in DeepSeek's table
DeepSeek's card also lists four other models. We reproduce the columns exactly as published; the other labs did not produce these figures for this comparison, so they are DeepSeek's reported numbers.
| Benchmark | Flash 1017 | K3 | GLM-5.3 | GPT-5.6 Sol | Opus 5 |
|---|---|---|---|---|---|
| HLE without tools | 42.6 | 43.5 | 42.0 (text-only) | 44.5 | 56.3 |
| Terminal-Bench 4.0 | 42.3 | 12.6 | 37.9 | 39.9 | 51.8 |
| DeepSWE v1.1 | 77.0 | 67.5 | 66.9 | 73.0 | 74.0 |
| ProgramBench | 29.1 | 17.5 | 19.0 | 23.0 | 37.0 |
| NL2Repo-Bench | 68.2 | 58.0 | 58.0 | 56.8 | 75.3 |
| Automation-Bench v1.0.6 | 60.5 | 46.7 | 48.8 | 45.8 | 50.3 |
| Agents' Last Exam | 35.4 | 27.6 | 28.5 | 26.7 | 28.6 |
| Apex Bench v1.1 | 72.1 | 47.8 | 56.6 | 54.3 | 65.8 |
| Chartography (tools) | 83.0 | 68.1 | not listed | 79.9 | 84.0 |
| BabyVision (tools) | 92.5 | 85.7 | not listed | 88.9 | 94.1 |
| ZeroBench-main (tools) | 51.0 | 41.0 | not listed | 53.0 | 52.0 |
Read it carefully. DeepSeek's model leads the table on DeepSWE, Automation-Bench, Agents' Last Exam and Apex Bench. Opus 5 leads on HLE, Terminal-Bench, ProgramBench, NL2Repo and the vision tests, except ZeroBench where GPT-5.6 Sol is ahead. A vendor chooses which rivals and which benchmarks to show, so a favorable table is expected. The relevant question is whether independent evaluators reproduce it.
Our earlier look at the budget tier, Haiku 5.5, GPT-6 Luna, DeepSeek, Gemini Flash and GLM, compared list prices and cache costs; the new release does not change the price comparison, only the quality side. For the lab behind Opus 5.5 and Haiku, see the Haiku 5.5 launch.
API pricing and the Search and Fetch release
A very long scroll unrolling with a marker, representing the 1M token context of DeepSeek V4.1 Flash 1017
The DeepSeek API pricing page now lists deepseek-flash as DeepSeek-V4.1-Flash-1017, next to deepseek-v4-pro (version 0813). The model has a 1M-token context, a maximum output of 384K, JSON output, tool calls, a Responses API, an Anthropic-format endpoint and vision.
| Price per 1M tokens | Off-peak | Peak |
|---|---|---|
| Input, cache hit | $0.003 | $0.006 |
| Input, cache miss | $0.15 | $0.30 |
| Output | $0.60 | $1.20 |
Off-peak rates are half of peak rates. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. For comparison, deepseek-v4-pro is $0.66 cache-miss input and $1.98 output off-peak.
Two pricing details are easy to miss:
- Legacy names redirect. The pricing page says deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired, and requests under those names are served by the 1017 model and billed at the Flash price. If you pinned the old name, your traffic already changed models.
- Search and Fetch. DeepSeek opened both APIs to everyone, at $0.75 per 1,000 successful requests each. When you use the API through DeepSeek Harness, Claude Code or Codex, server-side search and fetch calls are free for now; billing per request begins at 00:00 Beijing time on November 1, 2026.
The release lands two days after a five-hour API outage that DeepSeek refunded as credits. Reliability, not price, was the open question for many teams that week.
Can you run it locally?
The Hugging Face metadata lists about 763 billion parameters across BF16, F32, FP8 and int8 tensors, with 510 GB of safetensors. That is not a laptop model. A Hacker News commenter (jononor) praised the model as capable but said running it locally is very costly and wished for smaller "flash" models again, which matches the file size. The model card points to an inference folder for weight conversion and local inference, and recommends temperature 1.0 and top_p 1.0.
There is no Jinja chat template in this release. Instead, DeepSeek ships a reference Python encoder and a Rust-based toolkit called deepseek-recipe, with Python bindings, that converts Messages, Chat Completions and Responses requests into the V4 and V4.1 prompt format and parses output including thinking, tool calls and images. If you use your own serving stack, plan time for the prompt format, not just the weights.
If you want a no-cost way to test the previous generation in an agent harness first, see OpenCode's free access to DeepSeek V4.1 Flash, and Hao AI Lab's ORA robot harness, which ran the model zero-shot on robot tasks.
What developers are saying
The Hacker News thread had 30 points and about 10 comments in its first hours, so this is early reaction, not a test report.
- One commenter, freakynit, doubted the numbers, asking whether the benchmarks can really be real.
- Another, bashtoni, pointed out the gap between calling it "a minor update" and showing substantial gains; a second commenter, pmxi, made the same joke.
- jononor noted the model beating GLM 5.3 on the table, while saying local use is expensive.
Nobody in that thread had published hands-on results yet. We will add them in an update when they exist.
What to verify before you switch
- Run your own evals. Terminal-Bench and DeepSWE results depend on the harness. Use the same harness your team ships.
- Check output length and cost per task, not just price per token. DeepSeek claims shorter outputs; the saving shows up only if tokens per solved task fall. Measure it. Our DeepSeek V4 Flash cost-per-task look on ARC-AGI shows how to frame that.
- Watch peak hours. Doubling the price from 06:00 to 10:00 UTC on weekdays affects European and Asian mornings.
- Confirm what served you. If you used a legacy model name, your requests were silently rerouted.
- Data handling. The API is hosted by DeepSeek; if that is a compliance issue, the MIT weights are the alternative, at the hardware cost above.
Why it matters
Three things stand out. First, an MIT-licensed open model now claims parity with, or better than, closed models on several agent benchmarks, at a list price of $0.60 per million output tokens off-peak. If independent tests agree, open weights cover more of the agent workload. Second, DeepSeek keeps shipping Flash updates quickly, with the 1017 version following the earlier V4.1 Flash beta and API releases. Third, bundling a search and fetch service shows the company building a full agent stack rather than only a model, which fits its push for its own harness.
The caution is equally clear: the table is DeepSeek's, the rivals were chosen by DeepSeek, and the strongest claims are on benchmarks that many readers cannot run themselves. Wait for third-party results from groups that publish methods.
Related reading
- DeepSeek V4.1 Flash API beta and native multimodal architecture
- DeepSeek V4.1 Flash cuts KV cache HBM by about 75 percent
- Small models compared: Haiku 5.5, GPT-6 Luna, DeepSeek, Gemini Flash, GLM
- OpenCode makes DeepSeek V4.1 Flash free
- DeepSeek five-hour API outage and credit refund
- DeepSeek V4 Flash 0731 cost per task on ARC-AGI
- Hao AI Lab ORA with DeepSeek-V4.1-Flash
Details are accurate as of October 17, 2026. Benchmarks are DeepSeek's self-reported figures; prices come from the DeepSeek API docs and may change.