DeepSeek-V4.1-Flash: Specs, License & How to Run

DeepSeek-V4.1-Flash is a multimodal Mixture-of-Experts family from DeepSeek-AI. This page was first published from the original card on 2026-10-13. It was re-verified 2026-10-19 against the 1017 checkpoint, and re-checked 2026-10-21 against the same official card, LICENSE, news post, and DeepInfer.

Official hosts we opened on 2026-10-21: deepseek-ai/DeepSeek-V4.1-Flash-1017 (created 2026-10-17, not gated, MIT; 1,970 likes / 27,737 downloads), the LICENSE, DeepSeek’s 1017 news post, the original card deepseek-ai/DeepSeek-V4.1-Flash (created 2026-09-10, still online; 4,674 likes / 1,431,582 downloads), deepseek-ai/deepseek-recipe (1,598 stars, MIT), and deepseek-ai/DeepInfer (Apache-2.0, 13,104 stars). There is still no GitHub repository named deepseek-ai/DeepSeek-V4.1-Flash or …-1017 that we opened as a weights host.

This page is listed under ChatGPT alternatives and closed LLM APIs. Head-to-head with Z.ai’s MIT 320B/18B multimodal MoE: GLM-5.3-Flash vs DeepSeek-V4.1-Flash. Head-to-head with MiniMax’s October 2026 Community License MoE: MiniMax M3.1 Flash vs DeepSeek-V4.1-Flash. Versus Moonshot’s open 2.8T-class MoE: DeepSeek-V4.1-Flash vs Kimi K3. Versus Meta: Muse Spark Open vs DeepSeek-V4.1-Flash. Versus Ornith’s MIT 122B coding/agent MoE: Ornith-1.5-122B vs DeepSeek-V4.1-Flash. Serving: DeepInfer, vLLM, DeepInfer vs vLLM. Agent runtime: DeepSeek Harness.

1017 checkpoint update — what changed

Re-opened 2026-10-21. Only facts printed on the official 1017 model card, LICENSE, DeepSeek news post, or Hugging Face API. Adjectives such as “substantial” or “faster” stay vendor. We did not invent a latency number.

What the 1017 card states changed (copied labels):

• The README calls 1017 “a minor update to DeepSeek-V4.1-Flash.”
• It claims improvements across reasoning, coding, agentic tasks, and visual benchmarks.
• It claims the model reduces task execution time and output length, “enabling users to complete tasks faster and at lower cost” and “advancing the performance–cost Pareto frontier across multiple agent harnesses.” No numeric speed or token-length delta is printed next to that sentence.
• A new self-reported instruct table compares 1017 with original Flash (and with K3 / GLM-5.3 / GPT-5.6 Sol / Opus 5 — those extra columns are DeepSeek’s table, not a forkalt ranking).

What the 17 Oct 2026 news post states (API docs, not the weight README):

• “Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost.”
• “Now live on our website, app, and API.”
• Open-source hosts named: Hugging Face deepseek-ai/DeepSeek-V4.1-Flash-1017 and ModelScope deepseek-ai/DeepSeek-V4.1-Flash-1017.
• Search & Fetch API pricing on that post is API product news, not a statement about the MIT weight files.

What did not change (printed or unchanged on the 1017 card vs the original Flash card):

• MIT LICENSE; cardData.license: mit; not gated; pipeline tag image-text-to-text.
• No Jinja chat template. Same recommended sampling: temperature 1.0, top_p 1.0, 1M context, max_tokens ≥ 256K.
• Architecture size is not restated. We keep the original README’s 552B backbone / 8B prefill / 16B decode / 1M context / 196B Engram quotes.
• safetensors total remains 763,205,315,794. usedStorage is 510,296,920,807 bytes on 1017 vs 510,311,613,821 on the original card (HF API 2026-10-21). We do not invent a meaning for that 14.7 MB gap.

Field	1017	Original Flash card	Source
Host	deepseek-ai/DeepSeek-V4.1-Flash-1017	deepseek-ai/DeepSeek-V4.1-Flash	HF API
Created	2026-10-17T02:40:38Z	2026-09-10T02:17:58Z	HF createdAt
Last modified	2026-10-17T05:46:55Z	2026-10-16T09:27:28Z	HF API
Gated	false	false	HF API
Likes / downloads	1,970 / 27,737	4,674 / 1,431,582	HF API 2026-10-21
License	mit	mit	HF cardData + LICENSE
Pipeline tag	image-text-to-text	image-text-to-text	HF API
usedStorage	510,296,920,807 bytes	510,311,613,821 bytes	HF API

1017 README self-reported instruct table (full rows printed on the card we opened):

Benchmark	1017	Original Flash (on 1017 card)
HLE w/o tools	42.6 (44.0†)	36.8 (39.1†)
Terminal-Bench 4.0	42.3	31.2
DeepSWE v1.1	77.0	74.2
ProgramBench	29.1	20.3
NL2Repo-Bench	68.2	65.4
Automation-Bench v1.0.6	60.5	54.8
Agents’ Last Exam	35.4	31.8
Apex Bench v1.1	72.1	64.6
Chartography (w/tools)	83.0	78.9
BabyVision (w/tools)	92.5	89.6
ZeroBench-main (w/tools)	51.0	49.0

† denotes the text-only subset of HLE, per the 1017 README. Harness footnotes on that card: Claude Code for Terminal-Bench 4.0; Mini-SWE-V2 for DeepSWE v1.1. forkalt has not reproduced any row.

Arena.ai Code Arena WebDev (third-party, labeled)

The WebDev leaderboard with license=open-source, on the page we opened dated Oct 18, 2026 (900,155 votes, 58 open-source models), lists:

Rank (open)	Overall	Model	Score	Votes	License (board)
1	12	DeepSeek-V4.1-Flash-20261017 (Max)	1670	1,417	MIT
2	15	Kimi K3 (Max)	1655	19,426	Kimi K3 license
8	25	DeepSeek-V4.1-Flash-20260910 (Max)	1619	3,533	MIT

This is Arena.ai’s leaderboard, not a forkalt ranking. We do not treat “open-source #1” as a ChatGPT verdict or as an independent reproduction. The 1017 README extract we opened does not print that Arena rank; DeepSeek’s news post is the vendor announcement of the weights. See vs Kimi K3.

Architecture (original Flash README, still quoted)

Copied from the original card (checked first on 2026-10-13; architecture not contradicted by 1017): Causal Encoder-Decoder (CED), 40-layer Transformer (20-layer causal encoder + 20-layer decoder). CSA2 plus FP4 main KV caching is described as reducing global KV cache to 890 bytes per token. Vision encoder DeepSeek-ViT + two-layer MLP projector. Post-training: SFT → RL → on-policy distillation, with reasoning effort integer 1–100. Those sentences are upstream.

Hugging Face’s web UI has shown a “Model size 763B params” style line; 1017 safetensors.parameters sum to 763,205,315,794. We keep the README’s 552B backbone + 196B Engram quotes and do not invent a reconciliation.

How to run (from the 1017 README)

Recommended sampling copied from the 1017 card:

Parameter	Value
temperature	1.0
top_p	1.0
context_window	1M tokens
max_tokens	≥ 256K

No Jinja chat template. Prompt encoding: encoding/encoding.py on the card, production path deepseek-recipe (1,598 GitHub stars on 2026-10-21). Minimal inference: inference/ folder on the card.

The official serving framework published for this family is DeepInfer (Apache-2.0, 13,104 stars, 2026-10-21). Its README defaults to deepseek-ai/DeepSeek-V4.1-Flash-1017 and names NVIDIA SM100 / Huawei Ascend 950. Compare DeepInfer vs vLLM. Several other cards on this site still name vLLM (Apache-2.0, 94,459 stars, v0.32.0) or SGLang as a generic runtime; the 1017 Flash card itself points at inference/ and DeepSeek’s recipe, not a vLLM one-liner. forkalt has not executed those scripts.

Hosted chat.deepseek.com is a different artifact from downloading the MIT weights.

License

MIT, on the 1017 LICENSE file (copyright 2023 DeepSeek) and the README sentence “This repository and the model weights are licensed under the MIT License.” Same MIT tag on the original Flash card. Not legal advice.

What we did not invent

• Local VRAM, token throughput, and quantization recipes remain not independently measured.
• We do not invent a GitHub star count for a weights repo that 404s.
• We do not rank Flash against ChatGPT.
• Arena.ai ranks stay labeled third-party. DeepSeek table rows versus K3 / GLM-5.3 / GPT-5.6 Sol / Opus 5 stay self-reported.

Related: GLM-5.3-Flash — GLM-5.3-Flash vs DeepSeek-V4.1-Flash, MiniMax M3.1 Flash — MiniMax M3.1 Flash vs DeepSeek-V4.1-Flash, MiMo-V2.6-Flash, DeepSeek Harness, DeepInfer, vLLM, DeepInfer vs vLLM, Kimi K3, Reflection Beam, Muse Spark Open, Hy-Memory (uses DeepSeek V4.1 Flash as a named writer baseline in its own vendor table), Raon-1-mini. Self-hosted UIs: Open WebUI, LibreChat, Jan.