DeepSeek V4-Flash

DeepSeek V4-Flash is the smaller of the two large language models in the DeepSeek V4 family, a 284-billion-parameter Mixture of Experts model with 13 billion active parameters and a one-million-token context window, developed by the Hangzhou lab DeepSeek.^[19][20] It first shipped on April 24, 2026 as a preview, alongside the larger V4-Pro variant of the same DeepSeek V4 release. On July 31, 2026 DeepSeek put the official build, DeepSeek-V4-Flash-0731, into public beta on its API, keeping the same architecture, the same parameter count, and the same prices as the preview while reporting much higher scores on agent benchmarks.^[1][2] The 0731 release also added native support for OpenAI's Responses API wire format and a documented configuration for OpenAI Codex.^[1][5] Later the same day DeepSeek published the checkpoint as open weights under the MIT License in a new Hugging Face repository, deepseek-ai/DeepSeek-V4-Flash-0731, shipped with the DSpark speculative decoding module attached.^[29][31]

Scroll sideways for more →

Background

DeepSeek released the V4 series in preview on April 24, 2026, following DeepSeek-V3 in December 2024 and DeepSeek-R1 in January 2025, the two releases that made the lab internationally known. The series has two members: V4-Pro at 1.6 trillion total parameters with 49 billion activated, and V4-Flash at 284 billion total with 13 billion activated. Both support a one-million-token context and both were published as open weights under the MIT License.^[19][20]

At the same launch, DeepSeek pointed the two legacy API names, deepseek-chat and deepseek-reasoner, at V4-Flash in non-thinking and thinking mode respectively, and gave them a three-month retirement window ending July 24, 2026.^[1] That retirement took effect one week before the 0731 build shipped, so by the end of July the DeepSeek API exposed only two model ids: deepseek-v4-flash and deepseek-v4-pro.^[3]

Through the second quarter of 2026 the models on the API remained the April preview builds. DeepSeek's own July comparison chart labels them DeepSeek-V4-Flash-Preview and DeepSeek-V4-Pro-Preview.^[2]

What changed in the 0731 build

DeepSeek was unusually explicit about the scope of the update. According to the API change log, "DeepSeek-V4-Flash-0731 keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained."^[1] Nothing about the pre-trained backbone, the parameter count, the context length, or the price changed. The API calling method is unchanged as well: requests that already set model to deepseek-v4-flash picked up the new build automatically.^[1][9]

DeepSeek also scoped the update tightly in the other direction. The change log states that only the V4-Flash API was upgraded, and that the V4-Pro API and the models behind the DeepSeek app and website were left as they were, with the official V4-Pro release to "follow soon."^[1] Chinese-language coverage from IT之家 and TechNode reported the same split on the day.^[14][17]

Because the release is a post-training change rather than a new model, the practical effect for callers was a silent swap: the same model id returned different behavior from July 31 onward. That is worth knowing for anyone comparing evaluation runs across the boundary.

The model card published later that day with the weights adds one structural detail the change log left out. It describes DeepSeek-V4-Flash-0731 as "the official release of DeepSeek-V4-Flash, superseding the preview version, with substantially enhanced agentic capabilities," and states that "it has the same model structure as DeepSeek-V4-Flash-DSpark, i.e. it comes with a speculative decoding module attached."^[29] That module, DSpark, is a draft model bolted onto the checkpoint to speed up decoding rather than a change to the language model itself, and the shipped config.json carries its settings alongside the V4 architecture fields.^[29][30]

Benchmark results reported by DeepSeek

DeepSeek published nine agent-oriented scores in the change log and in a chart attached to its announcement post, and republished the same nine rows unchanged on the Hugging Face model card when the weights went up. The company's framing is that the new Flash build now beats the larger V4-Pro-Preview across the board, "despite its far smaller activated parameter count."^[1][2][29]

Scroll sideways for more →

All of these are vendor-reported figures. Two footnotes, carried identically on DeepSeek's launch chart and on the model card, matter for reading them. First, "for the Code Agent tasks among the public benchmarks above, DeepSeek-V4-Flash-0731 is evaluated with the minimal mode of DeepSeek Harness (to be released) as the agent framework, using the max reasoning effort level with temperature = 1.0, top_p = 0.95."^[1][2][29] A model's score on agent benchmarks depends heavily on the harness driving it, so results measured with an unreleased first-party scaffold are not directly comparable to leaderboard entries produced with a public one. The parenthetical "to be released" still stood on the model card that shipped with the open weights, which means the checkpoint being public does not by itself make these numbers reproducible.^[29] 36Kr noted that the announcement was the first time DeepSeek's in-house harness had appeared under an official name.^[16] Second, the dagger on the last two rows marks them as DeepSeek's own internal test sets: "DSBench-FullStack is an internal full-stack development test set; DSBench-Hard is an internal test set of difficult coding-agent problems." No outside party can reproduce those two rows at all.^[1][29]

What the benchmarks measure

Four of the public benchmarks have published documentation. Terminal-Bench 2.1 is an 89-task suite for coding agents working in a command-line environment, released on May 6, 2026 as a revision of Terminal-Bench 2.0 that repaired 28 tasks and added continuous validation of the task set.^[21] Toolathlon (the Tool Decathlon) covers 32 applications and 604 tools across 108 verified tasks that typically take around 20 turns, and its authors reported that the best model they tested reached 38.6 percent.^[22] CyberGym has 1,507 real vulnerabilities from 188 software projects and asks an agent to write a proof-of-concept test that reproduces each one from a text description plus source code.^[23] NL2Repo-Bench hands an agent a requirements document and an empty workspace and grades the Python library it produces against the original project's test suite, a setting in which the paper's authors found that even the strongest agents stayed below a 40 percent average pass rate.^[24] DeepSeek's rows are not directly comparable with the figures in those papers: its Toolathlon row is labelled "Toolathlon-Verified" rather than the full set, and every public code-agent run went through its own harness.

Independent evaluation

Artificial Analysis ran the model on the same day and published a breakdown that broadly supported DeepSeek's claim while trimming some of the numbers.^[10][11]

On the Artificial Analysis Intelligence Index v4.1, which aggregates nine evaluations (GDPval-AA v2, tau-cubed Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience and AA-LCR), DeepSeek V4 Flash 0731 in max reasoning effort scored 50, up from 40 for the April build and six points ahead of V4-Pro at 44.^[10][12] That put it level with Google's Gemini 3.6 Flash at 50 and one point behind GPT-5.6 Luna, Z.ai's GLM-5.2 and Meta's Muse Spark 1.1, all at 51. It remained seven points behind Kimi K3 from Moonshot AI at 57, which Artificial Analysis described as the open-weights frontier.^[10]

The agentic gains were the largest. Artificial Analysis measured GDPval-AA v2 at an Elo rating of 1559, up from 1189 for the preview, and its own Terminal-Bench 2.1 run at 79 percent, a 17-point rise but roughly four points below DeepSeek's reported 82.7. Its tau-cubed Banking score rose eight points to 31 percent. Every evaluation in the index improved: CritPt up nine points to 17 percent, SciCode up five to 50 percent, Humanity's Last Exam up five to 37 percent, AA-LCR up three to 66 percent and GPQA Diamond up one to 91 percent.^[10]

Two findings cut against a simple "it got smarter" reading. The AA-Omniscience result, an index score of -16 that Artificial Analysis called a 7-point improvement on the preview, came entirely from a lower hallucination rate rather than higher accuracy: the share of answers that were correct did not move, while the hallucination rate fell 12 points to 84 percent, which Artificial Analysis put in the same range as GPT-5.6 Terra at 85 percent.^[10] Separately, the model is verbose. It burned roughly 206 million output tokens as first reported on July 31 (Artificial Analysis's model page later showed 210 million) to complete the index, which is 12 percent fewer than the 234 million the April build needed but still far above the 62 million median across models.^[10][11]

The cost picture was the headline for Artificial Analysis. It placed V4 Flash 0731 on its Pareto frontier for intelligence against cost per task and calculated that, even after OpenAI cut GPT-5.6 Luna prices by 80 percent the day before, running the index on DeepSeek's first-party API cost about 60 percent less than on the comparably scoring OpenAI model. The firm attributed much of that to DeepSeek's roughly 98 percent cache-hit discount, against the 90 percent that most of the industry offers.^[10]

All of this was measured against DeepSeek's own hosted API rather than against the released checkpoint. As of August 1, 2026 no independent re-measurement of the open weights, run on someone else's hardware, had been published.

Responses API support and Codex compatibility

The 0731 release added a third wire protocol to the DeepSeek API, and for developers this may matter more than any of the benchmark rows. The Responses API is the request and response format OpenAI introduced for stateful, tool-using agentic work, and it is the protocol OpenAI's Codex clients speak. A provider that implements the same shapes can be dropped in behind a Codex install without patching the client. DeepSeek did exactly that at https://api.deepseek.com, and documents a ~/.codex/config.toml that sets wire_api = "responses" under a [model_providers.deepseek] block, plus a models.json catalog declaring the model's context window and reasoning levels to Codex.^[5][6] A one-line setup script is offered for macOS, Linux and Windows; DeepSeek says it backs up the existing configuration first and validates the files before writing.^[5] The same configuration serves the Codex CLI, the ChatGPT desktop app and the Codex extension for Visual Studio Code.^[5]

Responses support was no longer limited to Flash at the cutoff. V4-Pro gained native Responses support with its August 13 general-availability release, and the current catalog lists Responses support for deepseek-v4-flash, deepseek-v4-pro, and the experimental deepseek-v4-flash-vision-exp variant.^[1][3][6]

Compatibility limits

DeepSeek's implementation is a stateless subset. previous_response_id and conversation are not supported, store always returns false, and background, metadata, include, prompt and service_tier are not implemented. Requests that exceed the context window return HTTP 400 rather than truncating, because truncation is unsupported. prompt_cache_key and prompt_cache_retention are not accepted either, since DeepSeek manages context caching automatically. Unsupported parameters are silently ignored so that existing clients connect without modification. Among tools, function and web_search work, custom is accepted only for the apply_patch name that Codex needs, and file_search, code_interpreter, computer_use and MCP tools are ignored. The image rule now depends on the model id: Responses processes input_image only with deepseek-v4-flash-vision-exp, while the two non-vision ids replace it with placeholder text. Generic file inputs remain unsupported, although a Vision-Exp input_image can reference an uploaded image by file_id.^[6]

DeepSeek documents a separate integration for Anthropic's Claude Code through https://api.deepseek.com/anthropic, where its current name mapping sends Claude Opus names to V4-Pro and both Claude Haiku and Sonnet names to V4-Flash; callers can request Vision-Exp by its exact DeepSeek model id.^[8][50]

Experimental Vision-Exp API variant

On August 21, 2026, DeepSeek added deepseek-v4-flash-vision-exp to its API as an experimental multimodal variant, not as a separate general-availability release. The company said its text capabilities, including agent work, reasoning, and world knowledge, matched V4-Flash and described its multimodal agent results as close to Opus-4.8. Those are DeepSeek's claims. Contemporaneous coverage from The Paper reported the launch but did not supply separate tests.^[49][53][54] DeepSeek also released Harness 0.1.1 with support for the id; the DeepSeek Harness article covers that product's adapter and image-forwarding behavior.^[49]

The hosted catalog labels the model version DeepSeek-V4-Flash-Vision-Exp and advertises a one-million-token context, a 384K maximum output, thinking and non-thinking modes, JSON output, tool calls, Chat Prefix Completion, Responses API, and Anthropic API support. FIM completion is not supported.^[3] These are properties of DeepSeek's hosted route. On August 31, 2026, DeepSeek published a separate official checkpoint for Vision-Exp. Its configuration retains the 43-layer V4 language network and adds a 32-layer vision transformer with 1,024-dimensional embeddings, 16 heads, a patch size of 14, and at most 384 vision tokens. The reference implementation uses a two-layer aligner to map visual representations into the language model's 4,096-dimensional space.^[58][60][63] The vLLM recipe summarizes the full system as a 285B/13B MoE and estimates that the visual modules add about 0.5B parameters to the 284B backbone. DeepSeek's model card itself does not state a rounded total parameter count for the variant.^[61]

Input routes and limits

The documented interfaces accept text and image input and return text. DeepSeek does not document image generation for the variant. Chat Completions accepts images in user messages through an inline base64 data URL, a public HTTP(S) URL, a Files API file_id, or inline file_data; image content in system or assistant messages returns HTTP 400. The two non-vision model ids also return HTTP 400 when given images on this route. The Anthropic-compatible Messages route accepts base64, url, and file image sources, with anthropic-beta: files-api-2025-04-14 required for the last form. Responses accepts input_image in user or developer messages and in function or custom-tool outputs, but rejects images in system or assistant messages. On Responses, non-vision models substitute placeholder text for images, and generic file inputs remain unsupported.^[50][51][6]

Scroll sideways for more →

Images are billed as input tokens alongside text. DeepSeek says the Files API itself is free, but using an uploaded image for inference still incurs token charges. Uploads must finish within 10 minutes, and a stored filename may be at most 512 characters.^[49][50][51]

Vendor evaluation

DeepSeek published the following comparison. The values are vendor-reported results, not an independent leaderboard run.^[49][1]

Scroll sideways for more →

On the displayed numbers, Vision-Exp exceeds Opus-4.8 on DeepSWE, Agents' Last Exam, and ZeroBench, and trails it on the other eight rows. The chart therefore does not establish general parity. For text-based code-agent tasks in public suites, DeepSeek says it used DeepSeek Harness Minimal Mode with the output limit set to the maximum, top_p = 0.95, and temperature 1.0. It did not publish prompts, attempt counts, action or turn budgets, run logs, or confidence intervals for the chart. The comparison also mixes text and multimodal tasks, Pass@1 and Pass@5 metrics, and internal and public sets; the base model ignored multimodal elements in the ApexBench and Agents' Last Exam rows. These limits make the scores unsuitable for a blanket model ranking.^[49][1]

Artificial Analysis published an evaluation of DeepSeek V4 Flash Vision (Reasoning, Max Effort) on August 24. It reported an Intelligence Index v4.1.1 score of 51 for the hosted configuration.^[55][57] Its methodology describes the index as a primarily text-based, English-language suite, weighted 34 percent agents, 24 percent coding, 24 percent scientific reasoning, and 18 percent general capability; image-input evaluations are reported separately. The result is therefore independent evidence about a max-effort hosted configuration on mainly text tasks, not an independent validation of image understanding and not a reproduction of DeepSeek's multimodal benchmark table.^[56]

Artificial Analysis also reported 120.0 output tokens per second on DeepSeek's API, a weighted cost of $0.12 per index task, $235.89 in total evaluation cost, and 130 million output tokens versus a 61 million comparison median.^[55] These measurements are route- and suite-specific. Service conditions can affect speed; Artificial Analysis's cost method uses its task weights, provider token counts where available, and live measurements of typical cache-hit rates; output-token volume describes evaluation verbosity rather than quality.^[56]

The vLLM project later evaluated its pre-release open-weight serving implementation on all 1,000 OCRBench examples. The documented run used one GB200 NVL4 tray, tensor parallel size four, expert parallelism, an FP8 KV cache, a 32,768-token serving limit, and DSpark with three speculative tokens. vLLM reported 835 correct answers, no request errors, and 312.6 seconds for the run, about 3.2 samples per second.^[61][62] OCRBench's authors describe the original suite as 1,000 manually verified question-answer pairs across text recognition, scene-text visual question answering, document visual question answering, key information extraction, and handwritten mathematical expression recognition.^[65] The vLLM result tests OCR behavior and its serving implementation. It does not reproduce DeepSeek's hosted multimodal agent table or establish parity with the hosted API.

No separate checkpoint accompanied the August 21 API launch, and an exact search of DeepSeek's official Hugging Face organization still returned no Vision-Exp repository at the August 22 cutoff.^[49][52] That changed on August 31, when DeepSeek created deepseek-ai/DeepSeek-V4-Flash-Vision-Exp and uploaded an official 48-shard checkpoint under the MIT License.^[58][59][64] Vision-Exp has therefore been open weight since August 31, although DeepSeek continues to label the model experimental.^[49][58]

Architecture and specifications

Because the 0731 build reuses the preview's network unchanged, its architecture is the V4 architecture described in DeepSeek's technical report, "DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence," submitted to arXiv on April 26, 2026.^[20] The three changes the paper foregrounds are a hybrid attention design that pairs Compressed Sparse Attention with Heavily Compressed Attention, Manifold-Constrained Hyper-Connections (mHC) in place of ordinary residual connections, and the Muon optimizer. The paper reports that at a one-million-token context, V4-Pro needs 27 percent of the per-token inference FLOPs and 10 percent of the KV cache of DeepSeek V3.2.^[19][20] Both models were pre-trained on more than 32 trillion tokens. The DeepSeek V4 article covers the architecture in detail.

The config.json shipped with the 0731 weights puts numbers on parts of that description. It declares 43 hidden layers, a hidden size of 4,096, 64 attention heads with a head dimension of 512, LoRA-style query and output projections at rank 1,024, 256 routed experts plus one shared expert with six routed experts activated per token, an expert intermediate size of 2,048, a vocabulary of 129,280 tokens, and YaRN rope scaling by a factor of 16 from an original 65,536-token window up to the full 1,048,576. Expert tensors are marked "expert_dtype": "fp4" while the quantisation block declares FP8 E4M3 weights in 128 by 128 blocks with UE8M0 scales, which matches the "FP4 + FP8 Mixed" precision label on DeepSeek's V4 model cards: MoE expert parameters in FP4, most other parameters in FP8.^[30][35] Four dspark_ fields configure the attached draft module (block size 5, Markov rank 256, a dedicated noise token id, and target layer ids 40, 41 and 42), and num_nextn_predict_layers is set to 1.^[30]

Codex-facing metadata in DeepSeek's models.json fills in a few operational numbers that the model card does not: the declared context window is 1,048,576 tokens, Codex is told to treat 95 percent of that as effectively usable, parallel tool calls are always on, and the three reasoning levels are low ("Fast responses with lighter reasoning"), high ("Extra high reasoning depth for complex problems") and max ("Maximum reasoning depth for the hardest problems"), with high as the default.^[5]

Availability, pricing, and rate limits

V4-Flash is served from DeepSeek's first-party API at https://api.deepseek.com for OpenAI-format and Responses-format calls, and at https://api.deepseek.com/anthropic for Anthropic-format calls. The experimental Vision-Exp id uses the same two base URLs. DeepSeek changed the family to peak and off-peak billing on August 16, and Vision-Exp uses the same rates as Flash. The table below gives US dollars per million tokens.^[1][3][49]

Scroll sideways for more →

The Chinese-language pricing page lists the same schedule in yuan. Flash and Vision-Exp cost 0.05/0.10 yuan for cache-hit input, 1.5/3 yuan for cache-miss input, and 4.5/9 yuan for output; Pro costs 0.15/0.30, 4.5/9, and 13.5/27 yuan respectively, with off-peak first and peak second.^[4] This supersedes the April Pro promotion that had cut output from 24 yuan to 6 yuan per million tokens.^[15][1] The hosted catalog advertises a one-million-token context and 384K maximum output for all three ids.^[3] Concurrency is counted per account across API keys until a response completes, and excess requests receive HTTP 429. DeepSeek offers capacity expansion on request without an added expansion fee and supports a user_id parameter for per-tenant KV-cache, scheduling, and content-safety isolation.^[7] For a broader view of how these rates sit against other vendors, see LLM API pricing comparison.

Peak and off-peak pricing

The peak and off-peak schedule took effect at 16:00 UTC on August 16, 2026. Peak windows are 01:00 to 04:00 and 06:00 to 10:00 UTC, equivalent to 09:00 to 12:00 and 14:00 to 18:00 in Beijing; the lower rate applies at other times.^[1][3][4] Beginning at 00:00 Beijing time on Sunday, August 23, which was 16:00 UTC and noon EDT on August 22, DeepSeek's posted rule applies off-peak rates throughout Saturdays and Sundays in Beijing time.^[3][4]

At the posted peak rates, V4-Flash and Vision-Exp output cost 9 yuan per million tokens and V4-Pro output costs 27 yuan.^[3][4] The scheme reverses the direction of DeepSeek's earlier experiment with time-based pricing: from February 26, 2025 the company ran off-peak discounts instead, cutting DeepSeek-V3 by 50 percent and DeepSeek-R1 by 75 percent between 16:30 and 00:30 UTC, a program it wound down the following September.^[27] The plan was first communicated in an upgrade notice emailed to users on June 29, 2026, which tied the change to the V4 official release, then planned for mid July, and said users would be given 24 hours notice by email before any billing adjustment took effect.^[28]

Weights and licensing

The open weights landed the same day as the API beta, in a new repository rather than as a revision of the existing one. deepseek-ai/DeepSeek-V4-Flash-0731 was created on Hugging Face at 07:30 UTC on July 31, 2026, roughly half an hour after DeepSeek announced the API on X. The shards went up under a commit titled "Add files using upload-large-folder tool" at 09:54 UTC, and the commit titled "Release DeepSeek-V4-Flash-0731" landed at 12:02 UTC. A fourth commit, "add sglang cookbook to model card," followed at 03:07 UTC on August 1. The first three are attributed to the DeepSeek account msr2000; the August 1 model-card commit adds zjxia.^[31][32]

The repository holds 48 safetensors shards, with Hugging Face reporting 166,886,535,336 bytes of storage, about 167 GB. Alongside them sit config.json, generation_config.json, model.safetensors.index.json, the tokenizer files, LICENSE, README.md, and two directories. encoding/ carries Python scripts and test cases rather than a Jinja chat template, because this release ships no Jinja template at all: callers are expected to use the provided encode_messages and parse_message_from_completion_text helpers to turn OpenAI-format messages into input strings and to parse the model's output. inference/ holds a minimal local runner with a weight-conversion script.^[29][32][40] The license is MIT, the same as the preview.^[29]

The older deepseek-ai/DeepSeek-V4-Flash repository was left alone. It was last modified on June 22, 2026 and still opens with "We present a preview version of DeepSeek-V4 series," so anyone who had pinned that repository did not silently receive the new checkpoint.^[19][34] The preview weights remain there under the MIT License in FP4-plus-FP8 mixed precision, along with a base checkpoint, and are mirrored on ModelScope.^[19]

Artificial Analysis had written on launch day that DeepSeek "is expected to release the model's full weights in the coming weeks," and noted that on its GDPval-AA v2 result the model would then rank second among open-weights models behind Kimi K3.^[10] In the event the release commit landed the same day, and the firm's model page now lists the model as open weights under MIT with the checkpoint on Hugging Face.^[11]

Vision-Exp checkpoint

DeepSeek created the official Vision-Exp repository at 06:16 UTC on August 31, 2026. The large-file upload commit followed at 06:57 UTC, and the initial model-card commit landed at 07:57 UTC.^[64] Hugging Face records 48 safetensors shards and 167,819,616,863 bytes of repository storage, about 167.8 GB. The repository also contains the tokenizer, prompt encoder, image processor, model configuration, examples, and a minimal PyTorch inference implementation.^[58][59]

The model card licenses the repository under MIT and says the reference implementation covers the vision encoder and aligner, DFlash attention, the mixture-of-experts language network, Hyper-Connections, and the DSpark forward path.^[58] DeepSeek's inference README calls that code a readable reference implementation rather than a production serving engine.^[66]

Reading the parameter count

The figure Hugging Face displays beside a repository is derived automatically: its model API publishes a safetensors block that counts tensor elements per data type, and the sidebar shows that total.^[32] For the V4 checkpoints, whose experts are stored in packed FP4 and which ship a draft module in the same repository, that derived total is not the model's parameter count, and it is not even consistent between two repositories that DeepSeek says hold the same structure.

Scroll sideways for more →

Two things fall out of that table. The step from 46 shards and about 160 GB to 48 shards and about 167 GB is what attaching the DSpark draft module costs; the 0731 repository sits on the larger side of that step because it ships the module.^[32][33][34] And the two non-vision 48-shard repositories report the same file list, a byte-identical config.json and exactly the same total storage of 166,886,535,336 bytes, yet Hugging Face reports 165.3 billion tensor elements for one and 304.2 billion for the other. The entire gap sits in the data-type breakdown the API returns: the 0731 repository declares 296,352,743,424 INT8 elements and no E8M0 tensors, while the DSpark repository declares exactly half that many INT8 elements, 148,176,371,712, plus 9,261,408,000 E8M0 tensors, which the shared config.json identifies as quantisation scales ("scale_fmt": "ue8m0") rather than weights.^[30][32][33] Neither number can be read as a parameter count.

DeepSeek's own figure has not moved, and it is 284B total with 13B activated. The V4 model cards give that figure for DeepSeek-V4-Flash, the DSpark card repeats it for what DeepSeek says is the same checkpoint, and the July 31 change log says the 0731 build "keeps the same model architecture and size" as the preview.^{[1][19][35][36]} Third parties serving the released weights agree: vLLM's deployment recipe describes the 0731 variant as 284B total and 13B active, with MoE expert weights stored in FP4 and the remaining attention, normalisation and router parameters in FP8, and Baseten, which hosts the checkpoint, lists it as "284B total parameters and 13B activated per token."^[38][41] The 304.2B figure is therefore an accounting artifact of how the 0731 repository declares its FP4 expert elements and quantisation scales, not a parameter count and not the size of the DSpark module, which the shard and storage step puts nearer 7 billion. DeepSeek's own figure, corroborated by Artificial Analysis, vLLM and Baseten, remains 284B total and 13B activated per token. DeepSeek has published no statement addressing the discrepancy.

The two Hugging Face repository records show a narrower, internally consistent difference for Vision-Exp. Compared with V4-Flash-0731, its repository uses 930,881,442 more bytes and its tensor metadata contains 466,405,632 more elements. Almost the entire increase is BF16, with 466,393,088 additional BF16 elements and 12,544 additional F32 elements; all other datatype counts are unchanged.^[32][59] This is consistent with vLLM's estimate that the visual modules add about 0.5B parameters and with its rounded 285B label. It does not make Hugging Face's displayed 304.6B tensor total an authoritative parameter count for the model.^[61]

DSpark

DSpark is not a separate model. The model card for DeepSeek-V4-Pro-DSpark says so directly: the repository "is not a new model. It is the same checkpoint with an additional speculative decoding module attached." The Flash-DSpark card carries the identical note.^[35][36] Both point at DeepSpec, DeepSeek's MIT-licensed "full-stack codebase for training and evaluating draft models for speculative decoding," in which DSpark is one of three draft algorithms alongside DFlash and EAGLE-3 and is described as confidence-scheduled speculative decoding with semi-autoregressive generation.^[37] DeepSeek published DSpark variants of both V4 models on June 27, 2026, two months after the preview, and a set of small DSpark, DFlash and EAGLE-3 draft models trained on Qwen3 and Gemma 4 targets the following day.^[43] With the 0731 release the module stopped being an optional side repository and became part of the shipped checkpoint.^[29]

Running the weights

The model card documents two server stacks. Under vLLM, DSpark is switched on with a single flag, --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}', in a launch line that also sets an FP8 KV cache, a block size of 256, data-parallel size 4 with expert parallelism, the deep_gemm_mega_moe MoE backend and an FP4 indexer cache.^[29] Under SGLang, DSpark is enabled with --speculative-algorithm DSPARK and, the card notes, no separate --speculative-draft-model-path, alongside --moe-runner-backend flashinfer_mxfp4 and tensor-parallel size 4.^[29] The SGLang cookbook page for DeepSeek V4, added to the model card on August 1, generates launch commands across NVIDIA B200, B300, GB200 and GB300 configurations and warns that enabling speculative decoding makes SGLang reset --max-running-requests to 48.^[39] The vLLM recipe covers the same ground for NVIDIA hardware and adds AMD MI325X and MI355X configurations plus disaggregated prefill and decode setups.^[38]

Vision-Exp required a separate serving path at the September 1 cutoff. vLLM's multimodal implementation was still in open pull request #54566 rather than a stable release; the standard wheel routed the checkpoint to the text-only class and failed on its vision tensors. The project supplied the pinned vllm/vllm-openai:deepseekv4-flash-vision image until that support lands.^[61][62] Its verified configuration used tensor parallel size four with expert parallelism on one GB200 NVL4 tray. The run limited the model to 32,768 tokens, so it did not test the advertised one-million-token context, and vLLM did not report an end-to-end vision run on its listed H200, B200, or B300 alternatives.^[61] The project's September 1 announcement that vLLM "serves it now" referred to this pinned pre-release route, not to support in the stable wheel.^[67]

For running without a server, the inference/ directory documents converting the Hugging Face checkpoint with convert.py and launching with torchrun across a model-parallel group. To run the experts in FP8 instead of FP4, DeepSeek instructs users to remove "expert_dtype": "fp4" from config.json and pass --expert-dtype fp8 to the converter.^[40] For local deployment the card recommends temperature = 1.0 with top_p = 0.95 in agentic scenarios and top_p = 1.0 otherwise, and a maximum output length of 384K tokens at the high and max reasoning effort levels, matching the cap on the hosted API.^[29][3]

Third-party redistribution

Community conversions appeared within a day. A Hugging Face search on the model name returns GGUF builds from Unsloth, bartowski, ggml-org and lmstudio-community, NVFP4 and MXFP4 quantisations, MLX conversions for Apple silicon, and EXL3 builds at 2.0, 2.5 and 3.0 bits per weight.^[42] Baseten added the checkpoint to its model library and serves it behind an OpenAI-compatible endpoint.^[41] OpenRouter, which on August 1, 2026 was routing V4 Flash at discounted prices of $0.0896 per million input tokens and $0.1792 per million output tokens,^[25] lists the 0731 build as a separate entry, deepseek/deepseek-v4-flash-0731, whose catalog record carries a creation timestamp of July 31, 2026. As of August 5, 2026 that listing aggregated more than a dozen hosting providers, with routed prices starting at $0.09 per million input tokens and $0.18 per million output tokens at the cheapest host and DeepSeek's list rates of $0.14 and $0.28 at most others, including DeepSeek's own first-party endpoint; OpenRouter's original deepseek/deepseek-v4-flash slug is now labelled "DeepSeek V4 Flash 0423" and points at the April preview at the same $0.14 and $0.28 rates.^[46][47] Nous Research also offers the model through its Nous Portal subscription platform in partnership with GPU cloud provider Novita AI: the preview build was free for a limited period from May 14, 2026, and on August 2, 2026 Nous announced a seven-day 90 percent discount on the 0731 build.^[44][45] For readers comparing terms across labs, see open-weight license comparison and open source AI.

Reception and adoption

Coverage on launch day concentrated on the inversion of DeepSeek's own product ladder: a 13-billion-active-parameter model outscoring the lab's 49-billion-active flagship preview on agent work. The Chinese outlet 智东西 ran the story under a headline calling it "倒反天罡" (roughly, the natural order turned upside down) and argued that the release shows low price does not imply low capability.^[18] 36Kr reported DeepSeek's framing that for specific tasks, model scale is not the decisive factor once post-training optimization is applied properly.^[16] OfficeChai's write-up compared the reported scores against Opus 4.8 and against GPT-5.6 Sol, Claude Fable 5 and Claude Sonnet 5 list prices, all of which V4-Flash undercuts by a wide margin.^[13] Mainstream business coverage followed with the next wave of Chinese releases. In an August 4, 2026 story on China's model blitz, Bloomberg reported that a so-called DeepSeek death zone had emerged on a widely shared Artificial Analysis benchmark chart, for rivals that charge more for the same capability or deliver less for the same price. The story called V4-Flash "a breakthrough model in terms of pricing" and cited Artificial Analysis benchmark testing in which executing a complex real-world workload cost $0.03 with V4-Flash against $3.15 with Anthropic's Claude Fable 5.^[48]

Third-party coding agent products moved quickly. The agent tool Command Code announced support the same morning, saying the model was live on all of its plans with tool calls and cache repairs.^[26] Within a day of the weights going up the repository had gathered more than a thousand likes on Hugging Face and had been re-quantised into a dozen community formats.^[32][42]

Limitations

The base deepseek-v4-flash route remains text-only, while Vision-Exp accepts text and image input but documents text output only; DeepSeek does not document image generation for either route.^[3][6][50] The Responses implementation is stateless, so applications that rely on server-side conversation threading through previous_response_id need to manage history themselves.^[6] Verbosity is a real cost driver despite the low per-token price: Artificial Analysis measured roughly 206 million output tokens to complete its index, more than three times the median.^[10] The AA-Omniscience result shows an 84 percent hallucination rate, improved but still high, and factual accuracy did not move between builds.^[10]

DeepSeek released Harness 0.1.1 when it launched Vision-Exp, removing the earlier absence of a public harness.^[49] The vendor benchmark table is still not independently reproducible from the checkpoint alone because DeepSeek did not publish prompts, attempt counts, action or turn budgets, run logs, confidence intervals, or the internal DSBench-Hard test set.^[49][58] The vLLM OCRBench run uses a public suite and a documented configuration, but it measures OCR accuracy and serving behavior rather than the agent benchmarks in DeepSeek's launch chart.^[61][62][65]

Finally, the July 31 change-log entry still describes the base Flash API as public beta, while DeepSeek explicitly labels Vision-Exp experimental. V4-Pro reached general availability on August 13, and the current catalog lists Responses support for all three ids.^[1][3][49] Both open checkpoints are heavy for self-hosting: V4-Flash-0731 uses 48 shards and about 167 GB, while Vision-Exp uses 48 shards and about 168 GB.^[29][32][59] vLLM budgets about 202 GB of VRAM for the Vision-Exp weights before the KV cache, and its only published vision run used a GB200 NVL4 tray through pre-release support.^[61][62]

Significance

The 0731 release is a data point about where capability gains were coming from in mid-2026. DeepSeek changed no architecture, added no parameters and raised no prices, yet the model moved from 40 to 50 on an independent aggregate index and from 7.3 to 54.4 on one vendor-reported coding benchmark. DeepSeek attributes the change entirely to redone post-training, and reported its public code-agent scores using a first-party harness, so part of the measured gain belongs to the scaffold rather than the weights. The lab has not published a technical description of the 0731 post-training run.^[1]

Publishing the base checkpoint under MIT on the day of the API beta closed part of that gap without closing all of it. Anyone can run the model, fine-tune it, or serve it independently. DeepSeek later extended that availability to Vision-Exp with the August 31 MIT-licensed checkpoint and had released Harness 0.1.1 with the hosted variant on August 21.^{[29][49][58][59]} Those releases narrow the reproducibility gap, but they do not close it because the agent evaluations still lack complete run artifacts and one reported test set is internal.

The protocol work carries its own weight. By implementing OpenAI's Responses API and publishing a Codex configuration, DeepSeek made its model a drop-in backend for tooling that developers outside China already run, which lowers switching costs in a way a benchmark table does not. Paired with a roughly 97 percent cache-hit discount and prices below many comparably scoring closed models, that combination, rather than a frontier score, is the competitive position the release stakes out.^[3][5][10]

See also

• DeepSeek V4
• DeepSeek
• DeepSeek V3.2
• OpenAI Codex
• OpenAI Responses API
• Kimi K3
• Speculative decoding
• Open weights
• vLLM
• SGLang
• Context caching
• LLM API pricing comparison

References

1. ^Change Log (entries dated 2026-07-31, 2026-08-13, and 2026-08-21), DeepSeek API Docs
2. ^DeepSeek-V4-Flash Official API is now LIVE in public beta, @deepseek_ai on X, July 31, 2026
3. ^Models & Pricing, DeepSeek API Docs (accessed August 22, 2026)
4. ^模型 & 价格, DeepSeek API 文档 (accessed August 22, 2026)
5. ^Integrate with Codex, DeepSeek API Docs
6. ^Using the Responses API, DeepSeek API Docs
7. ^Rate Limit & Isolation, DeepSeek API Docs
8. ^Integrate with Claude Code, DeepSeek API Docs
9. ^Your First API Call, DeepSeek API Docs
10. ^DeepSeek V4 Flash 0731 scores 50 on the Artificial Analysis Intelligence Index, @ArtificialAnlys on X, July 31, 2026
11. ^DeepSeek V4 Flash: Intelligence, Performance & Price Analysis, Artificial Analysis (accessed August 1, 2026)
12. ^DeepSeek-v4-Flash-0731 Scores 50 On Artificial Analysis Intelligence Index, Creates Big Spike On Pareto Frontier, OfficeChai, July 31, 2026
13. ^DeepSeek Releases DeepSeek-V4-Flash-0731, Gives Opus 4.8-Level Performance At A Fraction Of The Price, OfficeChai, July 31, 2026
14. ^DeepSeek puts V4-Flash API into public beta, TechNode, July 31, 2026
15. ^DeepSeek-V4-Pro 开启 API 限时 2.5 折优惠，活动持续至 5 月 6 日前 (DeepSeek-V4-Pro opens limited-time 75 percent API discount through May 6), IT之家, April 25, 2026
16. ^DeepSeek V4 Official Version Released, 36Kr, July 31, 2026
17. ^DeepSeek-V4-Flash 正式版 API 上线公测，V4-Pro 正式版将"尽快"发布, IT之家, July 31, 2026
18. ^倒反天罡？DeepSeek-V4-Flash 正式版上线，干翻自家先发的 Pro 版, 智东西, July 31, 2026
19. ^deepseek-ai/DeepSeek-V4-Flash model card, Hugging Face
20. ^DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence, arXiv:2606.19348 (April 26, 2026)
21. ^Terminal-Bench 2.1, tbench.ai (May 6, 2026)
22. ^The Tool Decathlon: Benchmarking Language Agents for Diverse, Realistic, and Long-Horizon Task Execution, arXiv:2510.25726
23. ^CyberGym: Evaluating AI Agents' Real-World Cybersecurity Capabilities at Scale, arXiv:2506.02548
24. ^NL2Repo-Bench: Towards Long-Horizon Repository Generation Evaluation of Coding Agents, arXiv:2512.12730
25. ^DeepSeek: DeepSeek V4 Flash, OpenRouter (accessed August 1, 2026)
26. ^DeepSeek V4 Flash 0731 now live in Command Code, @CommandCodeAI on X, July 31, 2026
27. ^Off-Peak Discounts Alert, @deepseek_ai on X, February 26, 2025
28. ^DeepSeek V4正式版要来了，API高峰时段价格翻倍 (DeepSeek V4 official release coming, API prices double at peak hours), Jiemian News via Sina Technology, June 29, 2026
29. ^deepseek-ai/DeepSeek-V4-Flash-0731 model card, Hugging Face
30. ^DeepSeek-V4-Flash-0731 config.json, Hugging Face
31. ^DeepSeek-V4-Flash-0731 commit history, Hugging Face API
32. ^Model record for deepseek-ai/DeepSeek-V4-Flash-0731, Hugging Face API
33. ^Model record for deepseek-ai/DeepSeek-V4-Flash-DSpark, Hugging Face API
34. ^Model record for deepseek-ai/DeepSeek-V4-Flash, Hugging Face API
35. ^deepseek-ai/DeepSeek-V4-Pro-DSpark model card (README.md), Hugging Face
36. ^deepseek-ai/DeepSeek-V4-Flash-DSpark model card (README.md), Hugging Face
37. ^deepseek-ai/DeepSpec, GitHub
38. ^DeepSeek-V4-Flash-0731 deployment recipe, vLLM Recipes
39. ^DeepSeek-V4 cookbook, SGLang documentation
40. ^DeepSeek-V4-Flash-0731 inference/README.md, Hugging Face
41. ^DeepSeek-V4-Flash-0731, Baseten Model Library (accessed August 1, 2026)
42. ^Hugging Face model search for DeepSeek-V4-Flash-0731 (accessed August 1, 2026)
43. ^deepseek-ai models sorted by creation date, Hugging Face API (accessed August 1, 2026)
44. ^Deepseek V4 Flash is now free via Nous Portal for a limited time thanks to @novita_labs, @NousResearch on X, May 14, 2026
45. ^DeepSeek V4 Flash 0731 is now 90% off on Nous Portal, @NousResearch on X, August 2, 2026
46. ^DeepSeek: DeepSeek V4 Flash 0731, OpenRouter (accessed August 5, 2026)
47. ^Model list entry and per-provider endpoints for deepseek/deepseek-v4-flash-0731, OpenRouter API (accessed August 5, 2026)
48. ^China's AI Blitz Creates 'Death Zone' for Rival US Model Makers, Bloomberg News (Zheping Huang, Nectar Gan and Saritha Rai), August 4, 2026
49. ^DeepSeek-V4-Flash-Vision-Exp Release: Multimodal API Now Live, DeepSeek API Docs, August 21, 2026
50. ^Vision, DeepSeek API Docs (accessed August 22, 2026)
51. ^Files API, DeepSeek API Docs (accessed August 22, 2026)
52. ^Exact DeepSeek-V4-Flash-Vision-Exp search in the deepseek-ai Hugging Face organization (accessed August 22, 2026)
53. ^DeepSeek-V4-Flash-Vision-Exp is now live on the DeepSeek API Platform, @deepseek_ai on X, August 21, 2026
54. ^DeepSeek continues releasing new models: multimodal model launched, The Paper, August 21, 2026
55. ^DeepSeek V4 Flash Vision (Reasoning, Max Effort), Artificial Analysis, accessed August 25, 2026
56. ^Artificial Analysis Intelligence Benchmarking Methodology v4.1.1, accessed August 25, 2026
57. ^Artificial Analysis Changelog, August 24, 2026
58. ^deepseek-ai/DeepSeek-V4-Flash-Vision-Exp model card, Hugging Face
59. ^Model record for deepseek-ai/DeepSeek-V4-Flash-Vision-Exp, Hugging Face API (accessed September 1, 2026)
60. ^DeepSeek-V4-Flash-Vision-Exp config.json, Hugging Face
61. ^DeepSeek-V4-Flash-Vision-Exp deployment recipe, vLLM Recipes, September 1, 2026
62. ^Add DeepSeek-V4-Flash-Vision-Exp support, vLLM pull request #54566, opened August 31, 2026
63. ^DeepSeek-V4-Flash-Vision-Exp vision.py reference implementation, Hugging Face
64. ^DeepSeek-V4-Flash-Vision-Exp commit history, Hugging Face API
65. ^OCRBench: On the Hidden Mystery of OCR in Large Multimodal Models, arXiv:2305.07895
66. ^DeepSeek-V4-Flash-Vision-Exp minimal inference README, Hugging Face
67. ^DeepSeek-V4-Flash-Vision-Exp serving announcement, @vllm_project on X, September 1, 2026