DeepSeek-V4-Flash-0731

# Fix list: DeepSeek-V4-Flash-0731

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/deepseek-v4-flash-0731, the October 2026 research run, assessed 21 October 2026. Grade B, 62.4 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on DeepSeek-V4-Flash-0731: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Capability, 66 out of 100, up to 11.9 more on the total

Why it scored 66: From 18 independent benchmark results in 4 areas, each scored as its share of the way from the benchmark's floor to the best result any model has on the same leaderboard. By area, Long context 90, Coding 88, Reasoning and knowledge 71, Tool calling and agent tasks 45. Publisher-submitted results count at half weight, and numbers that appear only on the publisher's card don't count.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-capability):

How well the model does on the benchmarks that matter to an agent. We don't average a publisher's own numbers into a grade. A result counts in full when we ran it on our reference machine, or when an independent party ran it and published the method (the Berkeley Function Calling Leaderboard, Artificial Analysis, Epoch AI's Benchmarking Hub, the SWE-bench and Terminal-Bench leaderboards, the Open ASR Leaderboard and others that run models themselves). A result the publisher submitted to such a leaderboard counts at half weight. A number that appears only on the publisher's card is shown on the listing as the publisher's claim and doesn't count.

Each result is scored as its share of the way from the benchmark's floor (chance for multiple choice, zero for a pass rate) to the best result any model, open or closed, has on the same leaderboard and setting on the day we read it. A model that matches the best model anywhere scores 100 on that benchmark. For an error rate such as word error rate, lower is better and the scale runs the other way, and where the leaderboard gives no floor the rate is read against the best model's, so twice the best model's error rate scores 50. Speed as a multiple of real time (RTFx) runs from about 1 to over 20,000, so it's read on a log scale from real time to the fastest model. On the MTEB leaderboard, where most models score between 0.5 and 0.75, the floor is the median of every model the same tab lists, and its results count at half weight because MTEB doesn't say who ran them and most are submitted with the model. An arena rating (Elo) isn't a ratio, so its floor is 400 points below the best model on that arena, the gap at which the better model wins about ten games in eleven. Where a model has results at several settings (reasoning effort, say), its best one counts, because that's the setting an agent would run.

The results group into areas, each weighted within the category.

- Tool calling and agent tasks, 35 (BFCL, τ²-bench, Terminal-Bench and agent suites like them)
- Coding, 20 (SWE-bench Verified, LiveCodeBench, Aider Polyglot)
- Reasoning and knowledge, 20 (GPQA Diamond, MMLU-Pro, Humanity's Last Exam, composite indexes that publish their components)
- Instruction following, 15 (IFEval, IFBench)
- Long context, 10 (RULER, LongBench and long-context retrieval tests)

A speech model is read on its own areas, recognition accuracy (word error rate on the Open ASR Leaderboard's sets, 60) and speed (real-time factor, 40) for speech to text, and intelligibility and speaker similarity (Seed-TTS-eval, the Open TTS Leaderboard, 70) and speed (30) for text to speech. A publisher's own converted copy of the same weights (Qwen's -hf builds, say) counts as the same checkpoint. An embedding model is read on retrieval (60) and the other MTEB task types, classification, clustering, reranking and similarity among them (40), from the MTEB leaderboard. A vision model adds an image area (MMMU, ChartQA, DocVQA or OCRBench) at 20, and the others shrink to fit.

Capability is the weighted mean of the areas with results, and needs results in at least two areas (one for a speech or embedding model). With less than that it's pending for that listing, its weight is shared across the other four, and the total is capped at 61, so a model nobody has benchmarked independently can't reach B, and every model with results ranked B or better stays above it. The listing says so. Until 21 October 2026 the cap was 69, which put models nobody had benchmarked above ones that had been.

## 2. Risk, 50 out of 100, up to 10 more on the total

Why it scored 50: File safety 15, safetensors only, less 10 because prompts are built by running DeepSeek's encoding_dsv4.py from the repository and the card's serving commands pass --trust-remote-code. Provenance 15, the deepseek-ai organisation and a SHA-256 on every shard. Safety evaluation 0, none published. Agent guidance 0, the encoding README's notes about the parser and the reserved developer role are usage notes, read alike for all three DeepSeek listings. Licence risk 15, MIT with no deductions. Security programme 5, for the security@deepseek.com reporting address in DeepSeek's site footer, read as for DeepSeek-V3.2, with no security.txt and no public advisories.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-risk):

What could go wrong when you run it, and what the publisher has done about it. A higher score means less risk.

- 0 to 25, file safety. 25 for weights in safetensors or GGUF only, 15 when pickle files (.bin, .pt, .pth) sit beside them, 0 for pickle only. 10 off when the documented way to load it or to build its prompts runs code from the repository (`trust_remote_code`, or the publisher's scripts in place of a template).
- 0 to 15, provenance. 10 when the files come from the publisher's own organisation, the one its site or docs link to, and 5 for a checksum on every file.
- 0 to 20, safety evaluation. 20 when results are published and someone other than the publisher ran them or can rerun them, 10 for the publisher's own results, 0 for none.
- 0 to 15, guidance for agent use. 15 for documented prompt-injection guidance or an instruction hierarchy the model was trained to follow, 10 for a statement of what its trained refusals cover and the safeguards to add around it, 5 for general cautions, 0 for none.
- 0 to 15, licence risk. 15 to start, 5 off for each of these, a use policy or field-of-use limits beyond the law, terms the licensor can change or revoke, a revenue or user threshold, and a licence that changed without a dated note or whose link is broken. Not below 0.
- 0 to 10, a security programme. 5 for a way to report a flaw (security.txt, a disclosure policy or a security contact), 5 for advisories handled in public.

When our reference-machine runs start, measured robustness (prompt injection with AgentDojo, over-refusal with XSTest) joins Risk.

## 3. Reach, 70 out of 100, up to 7.5 more on the total

Why it scored 70: Memory 5, because the smallest build of 4-bit or better from a runtime's own library is LM Studio's MXFP4 GGUF at 156.4 GB, which with a fifth added needs 187.7 GB, more than our 107.4 GB machine and within 256 GB. DeepSeek's own release, 166.9 GB, sits in the same band. Runtimes 25, the cap, from vLLM v0.32.0, SGLang v0.5.22, llama.cpp release b11771, Transformers 5.20.0 and LM Studio's library, with Ollama not counted since its library entry is retired with no models pushed. Builds 10, DeepSeek ships one build, the FP8 and FP4 release itself. Terms 15, MIT with nothing added. Download 10, not gated. Hosted 5, DeepSeek's API moved to V4.1 Flash, but six inference providers on Hugging Face serve this repository live.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reach):

Whether an agent can run it, and where.

- 0 to 30, memory. The smallest build from the publisher or a runtime's own library that keeps 4-bit precision or better, at its file size plus a fifth for the runtime and a working context. 30 for 16 GB or less, 25 for 24 GB, 20 for 32 GB, 15 for 64 GB, 10 for what fits our reference machine (the GPU share of 128 GB of unified memory), 5 for 256 GB, 0 above.
- 0 to 25, runtimes. 5 for each of llama.cpp, Ollama, vLLM, SGLang, LM Studio and Hugging Face Transformers that runs the architecture in a released version, from the runtime's own code, library or documentation. A publisher's fork or an unreleased branch doesn't count. 25 at most.
- 0 to 15, builds from the publisher. 15 for quantised builds at two precisions or more (FP8, 4-bit, GGUF), 10 for one, 0 when only other people's quantisations exist.
- 0 to 15, terms of use. 15 for an OSI-approved licence with nothing added, 10 for a custom licence that lets anyone use it commercially for free below a threshold or under a use policy, 5 for research or non-commercial use only, 0 for no licence or unclear terms.
- 0 to 10, download. 10 when the files download with no account and no terms to accept, 5 for a click-through that needs an account, 0 when a person approves each request.
- 5, hosted too. The publisher's own API serves it, or at least two inference providers do, so an agent without the hardware can still call it.

When our reference-machine runs start, speed on that machine (tokens per second and time to first token at the build that fits) joins Reach.

## 4. Documentation and upkeep, 43 out of 100, up to 5.7 more on the total

Why it scored 43: Model card 8, the card and the DSpark card it names give architecture, parameters and context, with intended use, limitations, languages and cut-off missing. Training data 5, a general description in the report with no cut-off. Last change 15, a card edit on 1 August 2026, 81 days ago. Responsiveness 0, DeepSeek answered none of the 61 discussions. Version policy 10, the preview and DSpark copies keep their own repositories and the dated 0731 release says what it supersedes. File health 5, the files agree with each other, but reports of tool-call errors and looping and a pull request for a missing chat template wait without a reply from DeepSeek.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-upkeep):

- 0 to 20, the model card. Intended use, limitations, architecture, parameters, context length, languages and knowledge cut-off. 20 for all seven, 3 off for each that's missing.
- 0 to 20, training data. 20 for the data published or listed with sources and sizes, 15 for a data card that names the sources or the mix, the cut-off and the filtering, 10 for a general description with a cut-off, 5 without one, 0 for none.
- 0 to 20, the last change to the files, the template or the reference code. 20 within 30 days, 15 within 90, 10 within 180, 5 within a year, 0 older.
- 0 to 20, responsiveness. The publisher's replies to the 40 most recent questions and reports in the repository's discussions or issues. 20 for most, 10 for some, 5 for a few, 0 for none.
- 0 to 10, version policy. 10 for earlier versions kept in their own repositories with a dated note when one is replaced, 5 for earlier versions kept without notes or a first release, 0 when files were replaced in place without a note.
- 0 to 10, file health. 10 when config, tokenizer, template and card agree and no report of a broken file waits without a reply, 5 when one does, 0 when the files contradict the card.

## 5. Agent interface, 75 out of 100, up to 2.5 more on the total

Why it scored 75: Chat template 10, there is no Jinja template and the format is given as Python to run. Tool-calling format 10, DeepSeek documents how tools are declared and returned and ships its own parser, but only vLLM's recipe names a runtime parser, so the 25 line isn't met. Trained tool calling 20, several invoke elements in one call block. Structured output 15, DeepSeek's deepseek-recipe library documents JSON Schema response formats for V4 prompts. Reasoning control 10, chat mode with no thinking and three effort levels in thinking mode, with reasoning returned as reasoning_content. Context 10, 1M tokens trained.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-interface):

What a model and a developer need to talk to it. Speech, image and other models with no chat interface read the tool-calling and structured-output lines for what the model takes and returns (voices, timestamps, streaming, typed outputs), and the note says how.

- 0 to 20, the chat template. 20 for a template in the repository (Jinja, in the tokenizer config or its own file), 10 when the prompt format is only given as code you run.
- 0 to 25, the tool-calling format. 25 when the publisher documents how tools are declared and how calls come back, with the parser to use in a named runtime, 10 when only the template or a runtime's own recipe says it.
- 0 to 20, tool calling the model was trained for. 20 with several calls in one turn, 10 for one call at a time.
- 0 to 15, structured output. 15 when JSON or JSON Schema output is documented for the model with how to ask for it, 5 when the card claims it without saying how.
- 10, reasoning control. Thinking that can be turned off or budgeted per request, with where the reasoning text comes back.
- 0 to 10, context. 10 for 128K tokens or more trained, 5 for 32K to 127K. Context reached only by scaling the position encoding counts one band down.

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- Tool-calling format is 10 because DeepSeek names no runtime parser itself; the card links vLLM's recipe that names deepseek_v4, and the panel may read that link as the publisher naming it, which would make it 25
- Ollama's library page now says the entry was retired on 27 August 2026, where the earlier research recorded 25 September 2026; either way no local Ollama model exists, and whether Ollama's released engine loads deepseek4 wasn't checked from its code
- Unchecked, whether deepseek-recipe's V4 encoder matches the 0731 effort levels; its default names are DeepSeek-V4-Pro and DeepSeek-V4.1-Flash, though its V4 version covers the V4 family
- Counting deepseek-recipe 0.2.0 (20 October 2026) as the template for last change would raise it to 20; it's read here as a library-wide release, not a change to this model's files
- Version policy reads the 0731 card's supersedes statement, dated by the release name, as the note; the preview card itself carries no note
- Unchecked, whether Transformers 5.20.0 loads the attached DSpark draft module
- Agent guidance reads DeepSeek's parser and developer-role notes as usage notes, not cautions for agent use, so 0, as the pilot read DeepSeek V4.1 Flash and alike for all three DeepSeek listings. Reading them as general cautions would give all three 5
- Capability is computed by script and isn't scored here
- security@deepseek.com is a company-wide reporting address, not shown to cover the weights. It's credited as a way to report a flaw, as for DeepSeek-V3.2, since all three DeepSeek listings read the same site

## Weaknesses

- The smallest 4-bit build is 156.4 GB, so it needs about 188 GB and doesn't fit a 128 GB machine
- No Jinja template, so prompts come from running DeepSeek's encoding code or a runtime's own encoder
- No safety evaluation or agent-use guidance published, and only a company-wide security address
- DeepSeek hasn't replied to any of the 61 discussions, including reports of tool-call errors
- The card lacks intended use, limitations, languages and a knowledge cut-off

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- vllm serve deepseek-ai/DeepSeek-V4-Flash-0731 --trust-remote-code --kv-cache-dtype fp8 --block-size 256 --data-parallel-size 4 --enable-expert-parallel --moe-backend deep_gemm_mega_moe --attention-config '{"use_fp4_indexer_cache": true}' --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}' --tokenizer-mode deepseek_v4 --tool-call-parser deepseek_v4 --enable-auto-tool-choice --reasoning-parser deepseek_v4 (the card's 4x GB300 example plus the parser flags from vLLM's recipe)
- Build prompts with encoding_dsv4.py, deepseek-recipe's V4 encoder or vLLM's --tokenizer-mode deepseek_v4, since the repository has no Jinja template
- Tool calls come back in a DSML tool_calls block with invoke elements, string parameters carry string="true" and JSON values string="false"
- Set temperature 1.0 and top_p 0.95 for agent work, and allow up to 384K output tokens at reasoning effort high or max
- Without the hardware, call it through a Hugging Face inference provider (Together, DeepInfra, Novita, Baseten, Scaleway, Featherless), not DeepSeek's API, where deepseek-v4-flash now means V4.1 Flash
- LM Studio's library build is a 156.4 GB MXFP4 GGUF and LM Studio states 156 GB minimum system memory

## What the review panel asked for

- merge the template pull request
- stated memory needs
- ship a template with each release
- drop --trust-remote-code from the card
- a disclosure policy

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.