1. The Preview never counted — the official release is judged on post-training
On July 31, 2026, DeepSeek dropped a high-information-density change into its API docs update log: DeepSeek-V4-Flash-0731 official API is now in public beta. The calling convention does not change — the model string is still — but what is happening under the hood is, in equal measure, entirely different and entirely the same as the April preview.
What is completely different: the scores. DeepSeek's published benchmark table shows that the 0731 checkpoint now surpasses the V4-Pro preview across every Code Agent benchmark the company publishes: Terminal-Bench 2.1 82.7 (preview 61.8, Pro 72.1), Cybergym 76.7 (preview 38.7, Pro 52.7), DeepSWE 54.4 (preview only 7.3, Pro 12.8), Toolathlon-Verified 70.3 (preview 49.7, Pro 55.9), DSBench-FullStack 68.7, DSBench-Hard 59.6. This is a 13B-active lightweight model that, on the agent workflows most worth paying for, beat its own un-finalized flagship.
What is completely the same: the architecture. The Hugging Face model card is explicit: 0731 shares the exact weight topology with preview. Post-pretraining, 32T tokens of data stay frozen, the 284B total / 13B active MoE structure stays frozen, the 1M-token context window stays frozen, the CSA + HCA hybrid attention stays frozen, the mHC residual stays frozen, the Muon optimizer stays frozen. Without spending any new compute and without retraining a single token, DeepSeek used post-training to drag a model from "can write code but does not know how to be an Agent" to "can run Responses flows and let Codex call it". The open-source camp saw, for the first time, a clear signal of "lightweight iteration": after weights are released, a model can still keep changing.
2. What actually changed in 0731: three engineering pipelines
DeepSeek did not publicly disclose the post-training recipe, but cross-referencing release notes, the Hugging Face model card, and the vLLM launch command, at least three layers of infra-grade capability were stacked into 0731.
The first layer is DSpark speculative decoding shipped by default. The 0731 checkpoint has DeepSpec's DSpark module mounted as part of the released weights; serving 0731 with vLLM requires only a single flag. The DSpark paper reports 60–85% end-to-end throughput improvement over the MTP-1 baseline at matched aggregate throughput, and it is lossless. For Code Agent workloads — where every step regenerates dozens of tokens — that is direct latency gain.
The second layer is Responses API and Codex adaptation. 0731 is the first DeepSeek API release with native OpenAI Responses support, and it has been specifically adapted for Codex's tool-calling and context conventions. That means developers can switch an OpenAI client to the DeepSeek backend by changing the base_url and swapping the model string to — no adapter to maintain.
The third layer is a three-tier reasoning_effort switch. now supports , , and . DeepSeek's recommended sampling is temperature=1.0, top_p=0.95 (agentic) or 1.0 (other), and / modes recommend a maximum output length of 384K tokens. In other words, 0731 can switch between "fast-answer mode" and "deep-reasoning mode" like Claude Opus 4 does, and the switch toggles real trained-in capability, not a prompt-layer trick.
Stacked together, 0731's essence is: same pretraining, different post-training + different decode topology + different API shape, producing different agent behavior. This is a paradigm of operating an LLM as a continuously deployed service rather than a one-time model release.
3. Why the price matters: turning Agents into infrastructure-grade spend
0731's pricing lays the cards on the table:
- Input: 0.14 USD / million tokens (0.0028 on cache hit)
- Output: 0.28 USD / million tokens
For comparison: V4-Pro's output is 0.87 USD / million tokens — Flash costs one-third. The more direct comp is Claude Opus 4.8 at 5 USD / 25 USD and OpenAI Codex at 5 USD / 30 USD. 0731 puts an 82.7 Terminal-Bench Agent-tier score behind a price tag that is 5% of Opus 4.8. For the first time, an LLM call can be budgeted as infrastructure spend rather than SaaS spend.
4. The boundary of self-hosting: open source is not a silver bullet
0731's MIT license lets anyone pull the weights and self-host, but 304B parameters (including the DSpark draft module) sets a clear serving floor. DeepSeek's official vLLM recipe targets a single 4xGB300 node. Unsloth's unquantized dynamic GGUF lands around 103 GB at 3-bit, requiring approximately 110 GB combined RAM + VRAM. That makes self-hosting unfriendly for small and mid-size teams — a single H100 node with 80 GB of VRAM cannot run the 3-bit build, and a full 4xGB300 box is well beyond most annual cloud budgets.
But the door that matters most is not closed: open source + non-revocable MIT + official serving recipe means any major cloud or inference vendor can pull the weights and ship that day. Fireworks, in fact, listed a serverless endpoint for 0731 on July 31. This is another instance of DeepSeek ceding "distribution rights for a same-scale model" to the ecosystem at large — Claude Opus cannot be sold under its own name by third-party clouds; DeepSeek V4-Flash can.
5. Two warnings worth flagging
First, every score comes from DeepSeek Harness, which has not yet been released. DeepSeek's benchmark notes clearly state: Code Agent tasks were run with minimal mode, max effort, temperature=1.0, top_p=0.95. This setting is not directly comparable to third-party reruns. Until independent third-party verification of Cybergym and DeepSWE lands, production routing decisions should not rely solely on DeepSeek's self-reported numbers.
Second, legacy API strings are being deprecated. and were temporarily mapped to V4-Flash's two modes in the April 24 migration window, but DeepSeek has stated these two strings will be retired within three months — i.e. October 24, 2026. Any code still using these two strings should start migration planning now.
6. Personal take: once this cadence runs, the value is not in Flash itself
0731's real impact is not the score deltas — it is the cadence it demonstrates: after weights ship, a model can still iterate via post-training plus decode topology as a first-class loop. That cadence did not exist at V4-preview time in April; four months later DeepSeek has codified it with 0731.
Over the next 6–12 months this cadence will spread to three kinds of players:
- Other Chinese open-source labs: Kimi K3 has already open-sourced its full MoE training stack. Teams with comparable technical depth (StepFun, Zhipu GLM-5) will walk the same path when their preview turns official.
- Closed-source flagships: Anthropic's Opus / Sonnet already ship monthly small-step updates; after 0731, closed-source vendors' patch-style updates will be forced to be more frequent.
- Enterprise self-hosting: a model of this size — single-node 4xGB300 is the serving floor — is the sweet spot for enterprise on-prem Agent deployments. Since V4-Pro has not yet gone official, Flash is DeepSeek's most presentable card in the enterprise Agent market — and the one they will most want to defend.
Once V4-Pro official lands, it will follow the same post-training-driven path; and if Anthropic wants to hold its pricing on the Agent tier against V4-Pro plus V4-Flash, it will need to compete on either newer product shape or deeper tool-call reliability.
Bottom line: 0731 is not just a model upgrade. It is a paradigm signal that a model is no longer a snapshot. Open-source LLMs have finally written continuous deployment into their own operational loop.