DeepSeek V4.1-Flash-1017 Review (2026): Beats Opus 5 on Coding, Loses on Reasoning

A free MIT-licensed checkpoint that wins the coding-agent benchmark by three points and loses the hardest reasoning exam by thirteen. Both halves of that sentence are the product.

By AI Tool Scout · Published October 19, 2026 · 19 min read

What Is DeepSeek V4.1-Flash-1017?

DeepSeek-V4.1-Flash-1017 is the October 17, 2026 snapshot of DeepSeek's Flash model — the 1017 is a date, not a version number, and that naming tells you most of what the release is. It arrived on a Saturday, 37 days after the September 10 launch of V4.1-Flash, and it is a post-training refresh rather than a new architecture. DeepSeek's own model card calls it "a minor update to DeepSeek-V4.1-Flash." The company shipped it into the website, the mobile app and the API simultaneously; the API name is still deepseek-flash, the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp still resolve to it, and the price did not change by a cent. If you were already calling Flash, you have already been upgraded, whether you asked or not.

The architecture underneath is September's, and it is unusual enough to spell out. V4.1-Flash is a causal encoder–decoder rather than a plain decoder: 40 transformer layers split into a 20-layer encoder for prompt processing and a 20-layer decoder for generation, so the model activates 8B parameters per prompt token and 16B per output token. DeepSeek describes the backbone as 552B parameters, and the vLLM recipe is more precise: 552B of backbone plus roughly 196B of Engram memory, which is where the 763B total comes from — and the Hugging Face safetensors index for this checkpoint sums to 763,205,315,794 parameters, so both numbers are true of different things. The text stack has hidden size 5,120, 384 routed experts with 6 active per token plus one shared expert, and a 32-layer vision transformer in front of it. Attention is the interesting part: a small indexer inherited from V3.2-Exp scores compressed latents and keeps the best 512 per query after a candidate stage filters 2,048 blocks of eight, only four layers (2, 8, 14 and 20) compress their own KV, and every other layer reads that shared cache.

The consequence is a 1,048,576-token context whose KV cache costs about 890 bytes per token in FP4 — roughly a quarter of V4-Flash's footprint and, by DeepSeek's arithmetic, under a gigabyte for a full million-token session. Maximum output is 384,000 tokens. Input is text and images, output is text, and thinking mode is on by default with a non-thinking mode available. DeepSeek tuned the build for DeepSeek Harness v0.2.1-rc.1 or later and names Tencent CodeBuddy, WorkBuddy and OpenCode as supported surfaces. Weights are published on Hugging Face and ModelScope under the MIT licence as 48 safetensors shards totalling 510.3 GB.

Two things about the framing are worth noticing before the benchmark table. The first is the slogan. DeepSeek's post-training lead, Wu Yu, announced the release as "entering the era of democratized wish-driven coding," and the technical claim behind it is reinforcement learning with self-feedback — the model uses its own confidence scores and self-generated critiques as the reward signal, so no human raters and no separate reward model sit in the loop. DeepSeek says this is what improved end-to-end project work, from game builds to 3D modelling. The second is that this release is also a search infrastructure launch: DeepSeek opened its in-house search engine to outside developers as Search and Fetch APIs at $0.75 per 1,000 successful calls each, against roughly $5 per 1,000 requests for Brave Search — about 6.7× cheaper. Calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free until 16:00 UTC on October 31, 2026, after which they are billed too.

What Changed: the Pareto Claim, Examined

DeepSeek's pitch is a "Pareto improvement in performance and cost" — a stronger model that is also cheaper to run, with nothing traded away. Three claims sit inside that sentence, and they are not equally well supported.

• Scores went up on everything. DeepSeek reports 11 benchmarks and the 1017 build wins all 11, by 2.0 points on ZeroBench to 11.1 points on Terminal-Bench 4.0. That is a vendor claim on vendor harnesses, and the size of the Terminal-Bench jump — a 36% relative gain from a post-training refresh — is the single most suspicious number in the release. It is not impossible; it is large enough that you should wait for a third party to reproduce it.
• Tasks finish faster. The company says the model completes work in less time. There is no published wall-clock table, and time-to-first-token for a reasoning model is dominated by thinking length, so this claim is really a restatement of the third one.
• Outputs are shorter — and this is the mechanism. The technical report describes a length penalty applied during reinforcement learning that is conditioned on a reasoning-effort value between 1 and 100: the penalty grows with the number of reasoning tokens but decays exponentially as effort rises, so at low effort every extra token has to justify itself. The same report gives the trade curve — raising effort from 25 to 100 lifted average pass@1 across eight benchmarks from 67.1% to 76.3% while increasing output length 2.5×. In other words, the model did not learn to be terse in general; it learned to be terse per effort level, and the effort dial is now a direct cost control.

Here is the part DeepSeek cannot put in a launch post, because it requires someone else's harness. Artificial Analysis evaluated both builds on the same Intelligence Index, and the numbers line up:

That is a genuine Pareto move on measured output: fewer tokens, more score, less money, and it is the strongest evidence in this review that the release is what DeepSeek says it is. Two caveats keep it honest. First, the cost figure is an artifact of Artificial Analysis's own prompt mix and cache assumptions, not your workload. Second, shorter is relative: Artificial Analysis still classifies the model as very verbose at 210M output tokens against a field median of 140M, and because output bills at four times the off-peak input rate, length is still the largest single term in your invoice.

Benchmarks: Where It Wins, Where It Loses

DeepSeek's table reports 11 benchmarks against five rivals. We reproduce it in full because the shape matters more than any single cell — and because the two rows people will quote in opposite directions are both in it.

† Text-only subset of HLE. DeepSeek runs DeepSWE on the mini-swe-v2 harness and Terminal-Bench 4.0 on the Claude Code harness. All figures are vendor-reported as of October 17, 2026.

Where it wins

DeepSWE by 3.0 points over Opus 5 and 4.0 over GPT-5.6 Sol is the headline, and it is not a fluke of one row — the build also wins Automation-Bench by 10.2 points over Opus 5, Agents' Last Exam by 6.8, Apex Bench by 6.3, and it beats Kimi K3 on 10 of the 11 rows and GPT-5.6 Sol on nine. Against Kimi K3 the margin on Terminal-Bench is startling: 42.3 versus 12.6, which says more about how badly K3's scaffold handles that harness than about DeepSeek's terminal abilities. The most useful win for buyers is Automation-Bench at 60.5 — SaaS workflows driven through REST APIs, which is what most "AI agent" projects actually are — where it leads every model in the table by at least ten points.

Where it loses

Humanity's Last Exam, 42.6 against 56.3. A 13.7-point gap is not noise and it is not a harness artifact; it is the largest single gap in the table, it runs the wrong way, and it is on the benchmark least sensitive to scaffolding. Flash is also last on that row once you read the footnotes: GPT-5.6 Sol is 44.5, Kimi K3 is 43.5, GLM-5.3 is 42.0 on the same text-only subset. This model is not a reasoning model that happens to be cheap; it is an agent model that happens to be able to reason a little.

Terminal-Bench 4.0, 42.3 against 51.8. DeepSeek's own table has Opus 5 a full 9.5 points ahead on terminal use, which is awkward for a model whose selling point is agentic work. ProgramBench, 29.1 against 37.0 — a 7.9-point deficit on building programs from a specification — and NL2Repo-Bench, 68.2 against 75.3 on repository-scale generation are the same story from different angles: DeepSeek's gains are concentrated in iterative agent loops, not in one-shot construction of large artefacts.

Three disclosure problems in that table

• The rivals are last generation. DeepSeek compares against Claude Opus 5 and GPT-5.6 Sol, not the current Opus 5.5 or GPT-6 Sol. On Artificial Analysis's independent index, Opus 5.5 sits at 58 and GPT-6 Astra at 53, against DeepSeek's 45. Beating a retired sibling is a weaker claim than the headline implies, and nothing in the release acknowledges it.
• The harness flatters the vendor. Terminal-Bench 4.0 is run in the Claude Code harness, which is a legitimate choice and also the one most likely to produce a good number for the model being tested. Independent boards disagree about how much that matters: Vals measured the September build at 19.7 on Terminal-Bench 4.0 under mini-SWE-agent, against DeepSeek's own 31.2 on the same model. If that ratio transfers, the 42.3 is closer to the high twenties in neutral hands.
• Nobody has re-run DeepSWE yet. As of October 19, no independent DeepSWE score for the 1017 build exists; Artificial Analysis published an index score on October 18, which is a composite, not this benchmark. The 77.0 is therefore a claim, not a fact. Ridge, an independent leaderboard that tracks exactly this, kept every number on the row tagged to the September build and said it will only accept a new score when the board's own label carries 1017.

One more gap worth flagging for completeness: there is no GPQA, AIME or SWE-bench Verified row in this release. The 11 benchmarks are agentic, visual and automation tests. The nearest GPQA reference is the September build's vendor-reported 90.9 on GPQA Diamond, which was not re-run for 1017 and should not be quoted as this model's score. A table built around terminal use and SaaS automation, with the maths and science suites omitted, is a table built to answer one question — and it answers it well.

Pricing in 2026: Unchanged, Which Is the Point

DeepSeek did not touch the rate card, so this section is short and the comparison is where the value shows. Every figure below is approximate, in US dollars, checked on October 19, 2026 against DeepSeek's own pricing documentation, Artificial Analysis's model pages, Anthropic's published price list, and the provider listings we could verify. Confirm current rates on each vendor's page before you build — DeepSeek has changed its peak windows and model names more than once this year.

What the rate card looks like against the field

All prices approximate, in US dollars, checked October 19, 2026, and change often. DeepSeek rates are shown off-peak first, peak second. AA scores are on Intelligence Index v4.3.2 at each model's highest reasoning setting. Cost per index task is Artificial Analysis's weighted average across its own evaluation mix, which is not your workload.

The comparison that decides most purchases is $0.26 per index task against $1.82 for Opus 5.5, $1.99 for Gemini 4 Argon and $3.26 for GPT-6 Astra — a 7× to 12.5× spread for a model that is 13 points behind on the same index. Whether that is a bargain or a false economy depends entirely on what a failed task costs you. For a coding agent that runs 200 loops a day and a human reviews the diff, it is a bargain. For a single hard question whose answer goes straight to a client, it is not. Our LLM API cost guide covers the general levers — caching, batching, output length — that matter more than the headline rate.

Open Weights and What It Actually Takes to Serve

The MIT licence here is not the Kimi K3 licence with a friendlier name; it is plain MIT, with no revenue gate, no attribution trigger and no acceptable-use appendix. You can self-host, fine-tune, redistribute and sell, at any scale, without talking to DeepSeek. That is a real difference from most "open" frontier releases in 2026, and it is the strongest strategic argument for this model.

The hardware argument is better than we expected and worse than a workstation. The checkpoint is 510.3 GB across 48 shards — 475 GiB — and the vLLM recipe lists a VRAM minimum of 614 GB, a 1.2× headroom factor over the weights. That clears in exactly two shapes:

The KV cache is not the constraint, which is the whole point of the architecture: at 890 bytes per token, a one-million-token session holds under a gigabyte of global KV. Compare a conventional model of this size, where the same session would cost hundreds of gigabytes and cap your concurrency at one. DeepSeek's numbers are one quarter the HBM and one eighth the SSD of the previous Flash generation for the same sequence length, and expanding context 256× from 4K to 1M raises decode FLOPs by only about a quarter.

So: one rented eight-GPU node runs this model. At published 2026 neo-cloud rates an 8× H200 node costs roughly $30,900 a month, and at $0.60 per million output tokens off-peak the crossover against the API is about 50 billion output tokens a month — far above what almost any team sends. Self-host for data residency, latency, fine-tuning or air-gapping, not to save money. If you want a number for the Kimi comparison: K3 needs 1.56 TB of weights, eight B300s or MI355X and a production recommendation of 64+ accelerators; this needs 510 GB and one H200 node. That is the difference between "open weights" as a marketing phrase and open weights as an option. The hardware maths is in our Kimi K3 review.

Pros & Cons

5 Best DeepSeek V4.1-Flash-1017 Alternatives

Comparison Table

† Text-only subset. A dash means the figure is not published on a comparable run, not that the model scores zero. DeepSWE and HLE figures for DeepSeek, Kimi K3, GLM-5.3 and Claude Opus 5 come from DeepSeek's own October 17 table; nothing in that table has been independently reproduced for the 1017 build. Prices are approximate, checked October 19, 2026.

Which One Should You Choose?

• Choose V4.1-Flash-1017 if you run a coding or browser agent at volume. At ~$0.20 per DeepSWE task, ~230 tokens per second and a 98% cache discount, the economics are not close. Cap max_tokens, run it at a deliberate effort level, and put a human on the diff — this is a model you supervise, not one you trust with the final answer.
• Choose it for automation work specifically. Automation-Bench 60.5 is the highest in its table by ten points, and SaaS workflows driven through REST APIs are what most "agent" projects actually are. That score is the most relevant number in the release for business buyers.
• Do not choose it for research-grade reasoning. 42.6 on HLE with no tools, last in its own comparison set, and 13.7 points behind Opus 5. If your prompt looks like a physics problem rather than a pull request, pay the 7× and take Opus 5.5 or Gemini 4 Argon.
• Do not choose it for one-shot large artefacts. ProgramBench 29.1 against 37.0 and NL2Repo 68.2 against 75.3 say the same thing: it is strong in loops and weaker at generating a correct large system in a single pass.
• Self-host only if residency or latency forces it. One 8× H200 node at roughly $30,900 a month beats the API only above about 50 billion output tokens a month. Below that, the API is cheaper, better redundant and someone else's pager problem. If you do self-host, budget a day for prompt encoding — there is no Jinja template.
• Pick a US or EU host if the data is sensitive. Fireworks, Baseten, DeepInfra and Parasail all serve this model near list price, and the MIT licence means the weights are yours either way. Do not put client-confidential material through the free chat.
• Compare against V4-Pro before you assume the tier matters. Flash scores 45 on the independent index against V4-Pro's 36, costs a third as much on output, and is the only one with vision. V4-Pro is still on the rate card; it is not the model you want for agent work.
• Watch the next four weeks. Two things are pending: an independent DeepSWE run of the 77.0, and whatever Beijing does about open-weight publication rules. Neither changes what you can download today; both change what you should assume about November.

Frequently Asked Questions