DeepSeek V4.1-Flash-1017: MIT weights and a $0.75 search API

On 17 October 2026 DeepSeek published V4.1-Flash-1017 to Hugging Face under the MIT licence. It is a 763B-parameter checkpoint with a 1M-token context, and Terminal-Bench 4.0 moves from 31.2 to 42.3. DeepSeek also opened its Search and Fetch APIs to every developer at $0.75 per 1,000 calls per endpoint. The weights are the headline, but the search pricing is the part that will change how other vendors price their products.

What happened

The model card, as summarised by Munder Difflin, describes the release as “a minor update to DeepSeek-V4.1-Flash, delivering substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks.” It also says the update reduces task execution time and output length. On agent workloads, shorter outputs mean lower bills and less waiting.

On the numbers, AI Weekly, citing MarkTechPost, reports a DeepSWE v1.1 score of 77.0, up from 74.2. That puts it ahead of Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0) on that benchmark. The new checkpoint beats the previous Flash on all 11 published benchmarks, by margins of 2.0 to 11.1 points. It is not a clean sweep against the closed models, though. Opus 5 still leads on HLE, 56.3 against 42.6. By Munder Difflin’s count, Opus 5 tops six of DeepSeek’s eleven rows, including Terminal-Bench 4.0 and ProgramBench. All of these are DeepSeek’s own figures.

Hosted pricing is $0.15 per million input tokens, $0.60 per million output tokens and $0.003/M for cache reads. Munder Difflin notes these are off-peak rates and that peak hours cost double. The companion paper is titled “DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression”, so the engineering emphasis is on memory and serving cost rather than raw scale.

There is one operational detail that matters more than any benchmark. According to Tech Insider’s breakdown of the release, the checkpoint went live behind the existing deepseek-flash alias, with no code change required. AI Weekly adds that there was no version bump and no opt-out. If you call that endpoint in production, your model changed on the 17th whether you noticed or not.

The search API is the real signal

Our research desk first read the search pricing as $0.75 per million tokens. It is actually $0.75 per 1,000 calls. AI Weekly says “successful requests” and caps the endpoints at 5 QPS per account. Calls routed through server-side tools in DeepSeek Harness, Claude Code and Codex are free until billing starts on 31 October at 16:00 UTC.

The unit is different, but the point stands. DeepSeek is no longer just selling chat tokens. It is selling retrieval as a metered function with a published rate card. Tech Insider says the price undercuts most enterprise search APIs. AI Weekly calls it notably aggressive next to what larger cloud providers charge. Anyone selling RAG or agent search now has a public number to be compared against, whether or not their customers ever touch DeepSeek.

The price may not last. Tech Insider points out that DeepSeek has not documented how failed calls, timeouts or empty results are billed. It also argues the price should be treated as a launch promotion until it has survived a billing cycle. We agree.

What it means for builders

The case for owning a model rather than renting one keeps getting stronger, but “own” needs some care here. MIT means no usage restrictions. A 763B model is still data-centre hardware, as Munder Difflin points out. The repo ships no Jinja chat template, and its inference code describes itself as “a readable reference implementation rather than a production serving engine.” Self-hosting is realistic for teams with GPU capacity and serving expertise. Everyone else will reach this model through DeepSeek’s API or a third-party host. That is the same trade-off we discussed when Mistral Large 4 announced its open weights.

Cost per task is where it gets interesting. AI Weekly cites a 56-trial pilot by Kingy.ai. In that pilot, Flash-1017 completed the work for $0.0157, against $0.0577 for GPT-6.1 Sol and $0.0653 for Claude Sonnet 5.5. That is one small test, not a verdict. Still, it is a strong reason to route routine agent steps to a cheaper model and keep frontier models for the hard ones.

If weights are free and search is a line item, the model is not what makes your product defensible. What does is your proprietary data, the evaluations that tell you whether a model swap broke something, and how tightly the system fits the workflow it automates.

What to do now

• Re-run your evals on deepseek-flash today. The alias changed underneath you. If you need stability, self-host a pinned checkpoint, or add regression tests that catch behaviour drift.
• Price your retrieval per call. Work out what each search or fetch costs you now, then compare it with $0.75 per 1,000. Budget conservatively until the 31 October billing terms are clear.
• Test it as a router tier. Send bounded coding and agent sub-tasks to Flash-1017 and measure pass rate and cost against your current model on your own tasks, not on published benchmarks.
• Check data flows before sending anything sensitive. Hosted calls and self-hosted weights carry very different risk profiles. Our notes on data security and privacy in AI projects and the EU AI Act checklist for SMEs are a reasonable starting point.

The date to watch is 31 October. If the $0.75 rate holds once real agent traffic arrives, every retrieval vendor’s pricing page will be measured against it.

Frequently Asked Questions

Is the DeepSeek search API priced per token?

No. Search and Fetch are priced at $0.75 per 1,000 calls per endpoint, with billing starting on 31 October 2026 at 16:00 UTC. Before then, calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free for DeepSeek API users.

Do I need to change my code to use V4.1-Flash-1017?

No. It is served behind the existing deepseek-flash alias, so existing integrations picked it up automatically. That also means you should re-run your own evaluations, because the behaviour changed without a version bump.

Can I realistically self-host it?

The MIT licence allows it, but at 763B parameters it needs data-centre-class hardware and real serving work. The repo's inference code is described as a reference implementation rather than a production serving engine. Most teams will use it through an API.