What Is DeepSeek V4.1 Flash? 1M Context

1M-Token Context and 300+ Tokens/s: A First Look at DeepSeek V4.1 Flash

DeepSeek V4.1 Flash is now an officially launched DeepSeek model. Released on September 10, 2026, it combines a 552B-backbone Mixture-of-Experts design with a 1M-token context window, native visual understanding, and a new Causal Encoder-Decoder architecture. The official API model name is deepseek-flash, and the released weights are available under the MIT License. Community testing has also reported output speeds above 300 tokens per second in some configurations. An initial report described the earlier internal beta that preceded the public release. That report

Source: @0xLogicrw

Updated 2026-10-21: DeepSeek open-sourced its DeepInfer inference framework while the 1017 checkpoint claimed the top open model rank in Arena's Agent and WebDev leaderboards (see the Update below).

Key Takeaways

• DeepSeek V4.1 Flash officially launched on September 10, 2026. The model is live through the DeepSeek API, is rolling out in the DeepSeek app, and is available through supported partner services.
• The model has a 552B MoE backbone and a 1M-token context window. It uses a Causal Encoder-Decoder architecture with 8B parameters active during input processing and 16B during output generation.
• Community tests report roughly 190–427 tokens per second in API and local configurations, with results varying by workload, concurrency, hardware, and serving stack.
• Native multimodality is confirmed. The released model supports visual understanding of images alongside text.
• Off-peak API pricing is $0.003 per million cached input tokens, $0.15 per million uncached input tokens, and $0.60 per million output tokens. Peak rates are double those figures.
• The weights are publicly available under the MIT License, with deployment paths through Hugging Face, ModelScope, vLLM, SGLang, MLX-based community tooling, and other compatible software.
• The model is aimed at coding, agents, long-context work, visual understanding, web generation, and high-volume inference, although independent tests still show meaningful differences across tasks.

What Is DeepSeek V4.1 Flash?

DeepSeek V4.1 Flash is a fast, multimodal model from DeepSeek. It officially launched on September 10, 2026, as the smallest model in DeepSeek's new architecture family. DeepSeek positions it around four changes: a new model architecture, native visual understanding, higher capability, and lower inference cost.

The release status is no longer ambiguous. DeepSeek's API change log identifies V4.1 Flash as an official release, provides the deepseek-flash model name, publishes benchmark results, and documents its API pricing. The released checkpoint is also available publicly under the MIT License.

The temporary model ID deepseek-v4.1-flash-expires-on-0910 belonged to the internal beta that preceded the launch. It should not be used as the current API identifier. Developers now call the launched model with deepseek-flash.

This situation follows the pattern discussed in an earlier analysis of the DeepSeek V4 release, where naming signals and preview access provided useful clues without constituting a final model specification.

DeepSeek V4.1 Flash at a Glance

Specification	Current status
Developer	DeepSeek
Type	552B-backbone multimodal Mixture-of-Experts model
Release date	September 10, 2026
API model ID	deepseek-flash
Architecture	Causal Encoder-Decoder
Active parameters	8B during input processing; 16B during output generation
Modality	Text and native image understanding
Context window	1M tokens
Output speed	Community tests range from about 190 to 427 tokens per second in reported configurations
Off-peak pricing	$0.003/M cached input; $0.15/M uncached input; $0.60/M output
Peak pricing	$0.006/M cached input; $0.30/M uncached input; $1.20/M output
API availability	Live through the official DeepSeek API
App availability	Rolling out in the DeepSeek app
Partner availability	WorkBuddy, CodeBuddy, OpenCode Go, and ClinePass report support
Public weights	Available through Hugging Face and ModelScope
License	MIT
Production status	Launched and available; deployment performance still depends on infrastructure

How DeepSeek V4.1 Flash Works and What Makes It Different

The released model discloses substantially more about its architecture than the earlier beta did. DeepSeek V4.1 Flash is a 552B-backbone Mixture-of-Experts model built around a Causal Encoder-Decoder design. The technical materials describe a 40-layer Transformer organized as a 20-layer causal encoder followed by a 20-layer decoder.

The asymmetric design separates input processing from output generation. The model activates approximately 8B parameters per token during prefill and 16B parameters per token during decode. This is the central reason its total parameter count does not directly determine the amount of computation used for every token.

The model also uses a compressed KV-cache design. Compared with the previous generation, DeepSeek says V4.1 Flash requires one-quarter of the HBM and one-eighth of the SSD storage for its KV cache. That matters for long-context applications and agents that repeatedly reread the same repository, tools, instructions, or conversation history.

Technical commentary and released-model analysis also describe an approximately 196B Engram conditional-memory component. That figure describes additional lookup-style memory associated with the model and should be distinguished from the 552B backbone figure used in DeepSeek's official release positioning.

The most visible differentiator is Flash Throughput. Early API tests reported 420 tokens per second, while other reports measured approximately 333, 355, 375, 400, or 427 tokens per second in particular workloads. One local serving test on four DGX Sparks reported aggregate throughput above 200 tokens per second under a multi-stream configuration. These measurements vary with concurrency, prompt type, cache state, quantization, and serving software. The detailed speed test

A separate local deployment reported around 28–38 tokens per second for prose on three DGX Sparks, while other four-Spark tests reported higher throughput on code and structured tasks. Such variation is expected from different hardware and software configurations. It means that a single peak number should not be treated as a guaranteed service-level throughput.

DeepSeek V4.1 Flash is best described as a launched high-throughput model with unusually strong reported efficiency, not as a model with one universal speed figure across every workload.

Native image input is the second major differentiator. The official API release describes V4.1 Flash as having native multimodal support and visual understanding. The released model materials identify text and image input, while community tests have also exercised image-aware coding and visual tasks. The earlier beta included at least one non-vision configuration, but that limitation does not describe the launched API model.

The 1M-token context window is now an official specification rather than an early estimate. Yet context capacity does not automatically establish useful recall across an entire 1M-token prompt. No supplied benchmark isolates retrieval accuracy, long-context reasoning, or degradation at every prompt length.

What You Can Do With DeepSeek V4.1 Flash

The strongest early use case is fast coding and agentic development. Independent tests place V4.1 Flash at 74.2% on DeepSWE v1.1, compared with 62.7% for the previous V4 Pro score cited in the release materials. DeepSeek's official benchmark list also reports 90.9 on GPQA Diamond, 31.2 on Terminal-Bench 4.0, 88.1 on CyberGym, 54.8 on Automation-Bench, and 31.8 on Agents' Last Exam.

Those results are particularly relevant to coding agents that read large repositories, call tools, and repeat similar context. The asymmetric architecture and compressed KV cache are designed for this kind of input-heavy workflow.

Front-end generation is another visible application. A tester asked DeepSeek V4.1 Flash to create 100 HTML files with distinct visual designs and reported better web-design output than DeepSeek V4 Flash. Other demonstrations produced one-shot Three.js games, procedural 3D scenes, visual layouts, and a browser-based ocean simulation using hand-written GLSL. These are individual creative evaluations rather than standardized design benchmarks, but they point to practical use in rapid prototyping. The 100-file experiment

Other reported uses include:

• Browser and computer use: users connected the model to Codex-style and agent-harness workflows for browser-use, computer-use, and tool-calling tasks.
• Game development: community demonstrations include one-shot Three.js mini-golf, tower-defense, FPS, and open-world projects generated without external assets.
• Visual and 3D work: one tester reported a one-shot cloth simulation with textures and physics, while other demonstrations generated procedural environments and interactive scenes.
• Long-running coding agents: DeepSeek's cache design is intended to reduce the cost of repeated context. The official release specifically highlights cache-hit charges as an important part of agent economics.
• Large document tasks: the confirmed 1M-token capacity supports long repositories, transcripts, or document collections, subject to actual recall quality.
• High-volume inference: the low API rates and sparse active computation make the model suitable for workloads that need many repeated calls.

The model's reported speed makes it particularly suitable for interactive loops. Its behavior is not uniformly optimal across every task. Community users have reported that it can generate many tokens before delivering a response, and local tests show that prose can be slower than code or structured output.

How DeepSeek V4.1 Flash Compares

The following comparisons combine official benchmark results with early independent tests. They are directional rather than a single standardized leaderboard.

Model	Reported evidence	Comparison caveat
DeepSeek V4.1 Flash	74.2 DeepSWE v1.1; 90.9 GPQA Diamond; 31.2 Terminal-Bench 4.0; 88.1 CyberGym; 54.8 Automation-Bench	Official benchmark settings and independent task settings are not identical
DeepSeek V4 Pro	DeepSeek says V4.1 Flash surpasses V4 Pro across performance, cost, speed, and total task time; V4 Pro remains available through the API	Independent testing reports exceptions on some individual benchmarks
DeepSeek V4 Flash 0731	About 128 tokens per second in one cited API measurement; 54.4% DeepSWE in a comparison reported by the community	Different test and serving conditions
GLM 5.3 Flash	One independent tester judged GLM 5.3 Flash smarter in a particular comparison; Artificial Analysis scores reported so far place GLM 5.3 above V4.1 Flash overall	Single-account judgments and composite scores measure different things
Kimi K3	Artificial Analysis scores reported so far place Kimi K3 above V4.1 Flash overall, while a same-prompt task cost much more on Kimi K3	Cost, speed, and quality vary by workload
GPT-5.6 Sol and Claude Opus 5	DeepSeek's 74.2 DeepSWE score was reported above 73.0 for GPT-5.6 Sol, and V4.1 Flash leads both models on several coding and agent benchmarks	These are benchmark-specific comparisons, not a universal quality ranking

Artificial Analysis testing reported a 40-point Intelligence Index score for V4.1 Flash, above DeepSeek V4 Pro 0813 at 36 but below Kimi K3 at 44 and GLM 5.3 at 45. The same report put the average cost of an Intelligence Index task at about $0.27 for V4.1 Flash, compared with roughly $2 for Kimi K3 and GLM 5.3. It also reported approximately 197 tokens per second for V4.1 Flash.

The practical trade-off is capability per dollar and task type. DeepSeek V4.1 Flash performs particularly strongly on coding and agent benchmarks, while independent evaluations still show that other models can be preferable for some visual, prose, or general-purpose workloads.

Availability: How to Access DeepSeek V4.1 Flash

The current access path is the official DeepSeek API. Use the existing DeepSeek base URL and set the model name to:

deepseek-flash

The earlier temporary identifier, deepseek-v4.1-flash-expires-on-0910, was used during the internal API beta and is not the launched model's current public identifier. The API testing discussion documents that earlier transition period.

DeepSeek V4.1 Flash is also rolling out in the DeepSeek app. The app combined the earlier fast, expert, and image-understanding entry points into a single V4.1 Flash experience for everyday conversation, complex questions, and image understanding.

The released weights are available through Hugging Face and ModelScope under the MIT License. The surrounding ecosystem provides serving paths through vLLM and SGLang, while community conversions and runtimes provide MLX-based local deployment options for Apple Silicon. Local tests have demonstrated the model on three or four DGX Sparks and on a 256GB Mac Studio configuration using a 2-bit MLX quantization with Engram tables offloaded to SSD.

Partner access is also expanding. OpenCode Go, WorkBuddy including CodeBuddy, and ClinePass have reported support or availability. These services can apply their own quotas, pricing, regions, and implementation details, so their access terms should not be treated as identical to the official DeepSeek API.

DeepSeek's off-peak API prices are $0.003 per million cached input tokens, $0.15 per million uncached input tokens, and $0.60 per million output tokens. Peak rates are double those amounts. The new pricing took effect at 04:00 UTC on September 10, 2026.

What We Don't Know Yet

Several important details remain open:

• Benchmark comparability: DeepSeek's official scores and independent results do not use one shared evaluation protocol. The reporting available so far does not provide a complete, independently replicated comparison across every coding, agent, design, cybersecurity, and general-intelligence task.
• Real-world speed: Reports range from roughly 28 tokens per second for local prose generation to more than 200 aggregate tokens per second in multi-stream local tests and more than 300 tokens per second in some API tests. There is no single standardized hardware and latency table.
• Hardware requirements at scale: The weights are publicly available, but deployment depends heavily on quantization, SSD or memory offload, cache configuration, and serving software. The exact hardware needed for different context lengths and throughput targets remains workload-dependent.
• Long-context quality: The 1M-token context window is confirmed, but public sources do not establish retrieval accuracy or reasoning quality at every point within that window.
• V4.1 Pro timing: DeepSeek says V4 Pro requests will route to V4.1 Flash beginning at 04:00 UTC on September 14, 2026, until V4.1 Pro launches. The bundle does not provide a release date for V4.1 Pro.
• Safety and security performance: V4.1 Flash posts strong cybersecurity and agent results, but the available sources do not establish broader safety behavior, false-positive rates, or reliability across production security tasks.
• Operational guarantees: Third-party platforms report support, but quotas, regions, retention policies, and service-level behavior vary by provider.

The pricing evidence is now more concrete than it was during the beta. DeepSeek's published off-peak rates are $0.003 per million cached input tokens, $0.15 per million uncached input tokens, and $0.60 per million output tokens, with peak prices at twice those levels. Individual usage tests remain useful as examples of real workloads, but they do not replace the official rate card because caching, output volume, traffic period, and task design affect the final bill.

Frequently Asked Questions

What is DeepSeek V4.1 Flash?

DeepSeek V4.1 Flash is a launched 552B-backbone Mixture-of-Experts model from DeepSeek. Released on September 10, 2026, it uses a Causal Encoder-Decoder architecture, supports native visual understanding and a 1M-token context window, and activates 8B parameters during input processing and 16B during output generation.

Is DeepSeek V4.1 Flash released?

Yes. DeepSeek officially released DeepSeek V4.1 Flash on September 10, 2026. It is live on the DeepSeek API, its weights are available publicly under the MIT License, and it is rolling out across the DeepSeek app and supported partner services.

How do I access DeepSeek V4.1 Flash?

Use the official DeepSeek API with the model name deepseek-flash. The model is also rolling out in the DeepSeek app and is available through services including OpenCode Go, WorkBuddy, ClinePass, Hugging Face, and ModelScope. Local deployment is possible with the released weights and compatible inference software.

What is DeepSeek V4.1 Flash's context window?

DeepSeek V4.1 Flash has a 1M-token context window. DeepSeek lists this capacity in its release documentation and API materials.

How much does DeepSeek V4.1 Flash cost?

DeepSeek V4.1 Flash costs $0.15 per million uncached input tokens, $0.003 per million cached input tokens, and $0.60 per million output tokens during off-peak hours. Peak-hour prices are double: $0.30 for uncached input, $0.006 for cached input, and $1.20 for output.

Is DeepSeek V4.1 Flash multimodal?

Yes. DeepSeek V4.1 Flash has native multimodal visual understanding and supports image input alongside text. The official API release and model materials describe native vision support.

Is DeepSeek V4.1 Flash open source?

DeepSeek V4.1 Flash is available as open weights under the MIT License. The released checkpoint is available through Hugging Face and ModelScope, with local deployment paths provided by the surrounding inference ecosystem.

What to watch next is a broader set of independently reproducible evaluations covering speed, coding quality, image input, long-context recall, and local hardware efficiency. The next major product question is the release of V4.1 Pro, which DeepSeek says will end the temporary routing of V4 Pro requests to V4.1 Flash.

Update — 2026-09-10

New third-party benchmark results add a stronger quality signal than the earlier anecdotal comparisons. On OpenDesign Arena’s everyday-design tasks, DeepSeek V4.1 Flash reportedly scored 81.2—98% of GPT-6 Astra’s result—at $0.023 per task versus $1.61 for Astra and $3.66 for Claude Fable 5.1. These are benchmark-specific figures, not a general ranking, but they substantially strengthen the case for low-cost frontend prototyping. OpenDesign’s report

Early engineering tests also show that results depend heavily on the surrounding client. A seven-way production-code comparison reported a score of 76.69 when V4.1 Flash was paired with Claude Code, but 59.92 with DeepSeek Harness; the run took 49 minutes and recorded cache-hit rates above 99%. A separate cybersecurity benchmark reported improvements over the prior version, including 65.6% single-run CVE rediscovery, 84.4% pass@3, and 78.9% precision. The coding comparison and cybersecurity results remain independent reports rather than official evaluations.

Update — 2026-09-14

Fresh local-deployment reports further narrow the hardware picture. One tester reported running DeepSeek V4.1 Flash on four home GPUs without NVLink, measuring 21,400 tokens per second for an uncached 32K prefill, 280 tokens per second for a budget-50 decode, and up to 500 tokens per second in a single-stream decode configuration. These are community measurements from specific software and quantization settings, not guaranteed throughput. The four-GPU test

Smaller deployments also appear viable. A separate report measured about 32 tokens per second on one or two DGX Sparks, with 600K context and a 775K KV-cache configuration, while a community project documents a 2.9-bpw EXL3 setup for two DGX Sparks. The results strengthen the case that local operation does not necessarily require a large multi-node system, but they do not establish a universal minimum hardware configuration or performance profile. The DGX Spark report The two-Spark EXL3 project

Update — 2026-09-16

DeepSeek V4.1 Flash has picked up a new managed access path beyond the official DeepSeek API. Alibaba Cloud announced on September 15 that the model is now offered through its Token Plan, positioned for agentic workloads with the 1M-token context, fast inference, and high throughput already documented above, with plans starting at $6 a month. Alibaba's community writeup frames the offering as a setup and cost-optimization guide, walking through credit management and a three-step onboarding flow. See Alibaba Cloud's announcement and the accompanying community guide. As with the other partner services noted above, this gateway applies its own pricing, quotas, and terms, which differ from the official DeepSeek rate card.

Independent evaluation has also continued to accumulate. Artificial Analysis now lists DeepSeek V4.1 Flash at #6 of 113 models on its Intelligence Index, and Fireworks AI reported a DeepSWE agent run at $0.43 per task versus $6.52 for a comparable run on GPT-6 Astra at the same quality—driven by a 174:1 input-to-output token ratio and cache hits accounting for roughly 60% of the bill. See the Artificial Analysis model page and Fireworks AI's report. These remain third-party measurements rather than standardized results, but they reinforce the article's cost-efficiency framing.

Update — 2026-10-06

NVIDIA now publishes a DeepSeek V4.1 Flash NVFP4 checkpoint, adding another quantized distribution for local deployment. Community testers are comparing it with native weights and EXL3 conversions, although one early report cautioned that the checkpoint may be larger than the native weights, so the format alone does not establish a lower storage requirement. MiaAI_lab’s comparison note and Tech2Wild’s size observation

A separate uncensored FP8 checkpoint is also available from dealignai. A community report claims that the weights change refusal behavior, while reporting that higher reasoning settings can produce more refusals; these observations are not controlled safety evaluations or first-party DeepSeek findings. The report and its supporting evidence provide an initial signal for local safety-behavior testing, not a settled characterization.

Update — 2026-10-17

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17. It is a checkpoint update to V4.1 Flash, now live on DeepSeek's website, app and API. DeepSeek calls it "a Pareto improvement in performance and cost." The company says the model scores higher on benchmarks, finishes tasks faster and writes shorter outputs. It is tuned for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and DeepSeek recommends Harness v0.2.1-rc.1 or later. DeepSeek's announcement The weights are published on Hugging Face as deepseek-ai/DeepSeek-V4.1-Flash-1017 under the MIT License. OpenRouter is serving it with DeepSeek as the first provider. Users said they noticed a quiet backend change on the API several days before the announcement. The earlier observation

These are the reported changes against the original release figures cited above. DeepSWE v1.1 rises from 74.2% to 77.0%, and Terminal-Bench 4.0 rises from 31.2% to 42.3%. DeepSeek compares the new DeepSWE score with Claude Opus 5's 74.0%, not with the newer Opus 5.5, and Opus 5 still leads on Terminal-Bench at 51.8%. The benchmark summary One summary of the official cost curve puts the 77% DeepSWE setting at about $0.20 per task, against roughly $0.65 for GPT-6.1 Sol's best setting at about 75%. That breakdown One account reports that API pricing is unchanged from the original release. All of these benchmark figures come from DeepSeek's own results. None has been independently replicated yet.

Early hands-on reports are small but consistent with the efficiency claim. One same-prompt test at the highest reasoning setting took about 17 minutes and $0.46 with 1017, versus 20 minutes and $0.50 with the previous checkpoint. The side-by-side Another tester had 1017 review recent commits written by Opus 5.5 for about ¥0.3. It found one P0 bug and several P1 bugs, and Opus 5.5 later confirmed them. That test Some community accounts also report other details, none of them confirmed elsewhere yet:

• DeepSeek has opened search and fetch APIs to everyone at $0.75 per 1,000 calls for each endpoint. Report
• 1017 reaches decode speeds of about 244–300 tokens per second (one source).
• Before the update, the earlier checkpoint was briefly free inside OpenCode, with no stated end date. Report

DeepSeek has not published a separate parameter count or context window for 1017.

Update — 2026-10-18

The first independent results for DeepSeek-V4.1-Flash-1017 are now available. On browser-use's long-horizon browser benchmark, 1017 scored 75/100. That is the same as Claude Opus 5.5, at $0.22 per task versus $2.35, or about 19x cheaper. browser-use still names GPT-6.1 Sol as the top model on that test. browser-use's result

A harness comparison by one tester gave 1017 a maximum of 39.1 over seven runs with the Claude Code harness and 42.7 with DeepSeek Harness. DeepSeek's harness added 2–3 points on average but cost about 80% more. On the same test, the maximum scores for other models were 45 for GPT-6 Astra, 42 for Muse Wonder Full 1.0 and 41.7 for Opus 5.5. The harness test

This differs from the September comparison above, where the Claude Code client scored well ahead of DeepSeek Harness. The checkpoint, task set and harness versions are all different, so the two results cannot be compared directly.

Same-prompt tests against frontier models show large cost gaps but mixed quality:

• On one coding task, Opus 5.5 took 20 minutes and about $6, while 1017 took about an hour and $0.23. The tester still judged Opus the better result.
• On another task, 1017 finished in 3 minutes for $0.78, while GPT-6.1 Sol took 5 minutes and $6.10.

Both tests were run by the same person. Reasoning-effort and harness settings vary from test to test, so these comparisons are not like-for-like.

Local deployment options are also growing:

• Unsloth quants: Unsloth has begun publishing quants of 1017. The first two are UD-Q4_K_XL at 501GB and UD-Q8_K_XL at 508GB, so the 4-bit build is barely smaller than the 8-bit one. The Unsloth report
• Four DGX Sparks: One tester reported 89.8 tokens per second for a single user (48.5 on vLLM) and 202.9 tokens per second across six users, using a 3.5-bpw EXL3 build.
• Mac hardware: Another tester ran it on two M3 Ultra 512GB machines with an mxfp4 build.
• ModelScope: ModelScope is now hosting the 1017 checkpoint.

A separate post repeats the earlier claim that DeepSeek has commercialized its internal search engine as a Search & Fetch API alongside the release. That summary DeepSeek's official release post still does not mention it.

Promotional accounts are circulating specs for 1017, including "763B FP8" uncensored variants. None of these figures has come from DeepSeek.

Update — 2026-10-21

DeepSeek released DeepInfer under the Apache 2.0 license, open-sourcing the production inference engine built to power V4.1 Flash at scale—an infrastructure release noted by Thom Wolf alongside DeepSeek's earlier FlashMLA and DeepEP releases. The 1017 checkpoint also entered the Arena leaderboards as DeepSeek-V4.1-Flash-20261017 (Max), ranking as the #1 open model in both Agent Arena and Code Arena: WebDev.

Comprehensive multi-client evaluations show substantial sensitivity to agent tooling. Across an 89-run test covering nine agent harnesses, Pawel Huryn recorded an approximately 10-point spread in task success: scores ranged from 31.8 on Kilo Code and 36.4 on Claude Code up to 38.6 on OpenCode 2.0, 39.2 on DeepSeek Harness, and 40.4 on Pi and Grok Build CLI. The variance reinforces that agentic performance depends heavily on the surrounding runtime rather than raw weights alone.

Local hardware barriers have also lowered. Unsloth AI released a 1-bit quantization that cuts model storage from 508GB down to 250GB while reportedly retaining ~88.2% accuracy, allowing local operation on machines with 128GB+ RAM. In parallel, OpenCode updated its Go and Free tiers to the 1017 checkpoint, and community repositories such as MiaAI-Lab documented abliterated TensorFold deployments for multi-accelerator local clusters.

Try related models on Kie.ai: DeepSeek-V4.1-Flash, Claude Haiku 5.5 and Gemini 3.8 Flash.