Skip to content
Newsroom
Models 2h ago by Rajat Jain

DeepSeek V4-Flash 0731 vs Max: Same Model, Two Labels

DeepSeek's API has two models, not three — 'Max' is an effort label on deepseek-v4-flash (0731). What the July 31 retrain changed, and why comparisons break.

DeepSeek V4-Flash 0731 vs Max: Same Model, Two Labels

Search “deepseek v4 flash 0731 vs max” and you will find trackers comparing the July 31 release against something called “Max” — with price and benchmark deltas. The official documentation settles it: there is no “V4-Flash-Max” model. DeepSeek’s API has exactly two models — deepseek-v4-flash (version DeepSeek-V4-Flash-0731 since July 31) and deepseek-v4-pro — and “Max” is what trackers call the max-effort configuration of the same model. This article is the disambiguation page for the open-weight cluster on The AI Breakout’s coding hub, and it reconciles the pair in the same way as our V4-Flash vs Qwen3.8-Max and Kimi K3 vs V4-Flash pieces: official numbers only — the standard extended by Qwen3.8-Max vs Kimi K3: The Numbers Beat (shared official DeepSWE board rows) and the Open-Weight LLM API Pricing 2026 rate-card table.

The two models

V4-Flash (0731)V4-Pro
Total / active params284B / 13B1.6T / 49B
Context / max output1M / 384K1M / 384K
StatusPublic beta, Jul 31, 2026Preview since Apr 24, 2026
Price (per 1M tokens)$0.14 in / $0.28 out$0.435 in / $0.87 out
Cache hit$0.0028$0.003625
Concurrency limit2,500500

That is the complete official model list, per the models & pricing page. “Max” appears nowhere in it. Where it does appear is in the evaluation lane: DeepSeek’s benchmark note describes 0731 being tested “with the max effort level” — an inference configuration, not a product. Trackers inherited the word: Artificial Analysis labels runs “Reasoning, Max Effort”; the LLM-stats board calls the same configuration “Flash-Max.” The label is a setting of one model, and the “vs Max” question is really “which effort level should I pay for, on a model that changed in July.”

What the 0731 retrain changed

The official changelog is unambiguous on the identity question: “DeepSeek-V4-Flash-0731 keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained.” A 284B/13B MoE, 1M context, MIT weights — unchanged. What moved was agent capability, on the official card:

  • Terminal-Bench: 56.9 on TB 2.0 for the April preview era, 82.7 on TB 2.1 for 0731 — note the benchmark version also changed, which is precisely the kind of detail tracker columns lose
  • DeepSWE: 7.3 → 54.4 (a 7.4x jump, the largest single official delta in the release)
  • NL2Repo: 39.4 → 54.2
  • Cybergym: 38.7 → 76.7
  • Toolathlon (verified): 70.3 · Automation Bench: 25.1 · DSBench-FullStack: 68.7 · DSBench-Hard: 59.6

The same release made deepseek-v4-flash the API string for the new build (existing integrations work unchanged), added native Responses API support and Codex adaptation, and explicitly left V4-Pro and the APP/WEB models untouched. DeepSeek also states the retrain’s agent results “far exceed” V4-Pro-Preview — a 13B-active model out-scoring its 49B-active flagship sibling on agent work, purely from post-training.

What other coverage missed: the two facts that dissolve the “vs Max” question — DeepSeek’s own measurement note says the retrain was evaluated with the max effort level, so the “Max” in every tracker table is an effort setting of this one model, and Hugging Face hosts no “V4-Flash-Max” repository at all (the org’s V4 line is Flash-0731, Flash, and Pro). Since the architecture is identical, a “Max beats 0731” headline is unverifiable in principle: there is no second model to beat it with.

Why comparisons break

Three traps produce the wild price and score spreads between outlets:

  1. Label conflation. BenchLM’s live profile literally titled “DeepSeek V4 Flash (Max)” is pointed at the July 31 release at $0.14/$0.28 — while Artificial Analysis applies “Max Effort” to the April preview’s effort tier. Same label, two referents, opposite release dates. Practitioners hit the same wall: the day of the release, the most-upvoted question on the developer tracker was whether 0731 was even live on the API yet — that is how little the naming tells you about what you are actually running.
  2. Cache-hit vs cache-miss. Official discount pricing is $0.0028 per 1M cached input tokens — a 50x spread against the $0.14 miss price. Trackers blend these differently, and any “price per token” column is only comparable if the cache ratio is stated.
  3. Effort-tier verbosity. Max-effort runs of the same model emit roughly 5x the output tokens of standard effort on reasoning workloads, per independent token accounting — so per-task cost moves an order of magnitude while the quoted per-token price does not.

The combination — a Max-shaped label, 50x cache spread, and effort-dependent verbosity — is why one outlet can call V4-Flash “a rounding error per task” and another can call it “expensive,” using the same official price sheet.

Which one should you use

  • Default high-volume agentic work: deepseek-v4-flash (the 0731 build) at standard effort — $0.14/$0.28, 2,500 concurrent calls, the new agent benchmark profile. This is the price-quality answer for sweeps, edits, and test loops.
  • Hard, long-horizon loops: the same model with thinking enabled at high/max effort where the output-token premium (5x) is justified by the task.
  • Flagship planning and deep reasoning: deepseek-v4-pro at $0.435/$0.87 if the ~5-point intelligence-index gap shows up in your workload — most teams never need it for coding.
  • Local: the 0731 weights are MIT-licensed on Hugging Face, ~160GB in full quantized form (GGUF builds already published) — see our V4-Flash cost deep-dive for the hardware math.

One practical note from the official pricing page: DeepSeek has announced a significant API price increase is expected in the near future — if 0731 economics are your baseline, the current rate card is the low-water mark to plan against.

What to watch

  • The official V4-Pro release — “will follow soon” per the changelog, with Responses API support slated for early August; the Pro question will get a clean answer then.
  • The announced price increase — the first since the April lows, and the first real stress test of the “cheap open-weight floor.”
  • Next retrains — the 0731 pattern (post-training over scaling) says the next Flash checkpoint, not a new architecture, is the likely event.

Official source

First published August 12, 2026. All model facts from the official DeepSeek API documentation and model cards; the preview-era baseline numbers are the ones DeepSeek published with the release.

#Models #DeepSeek #Open Source