DeepSeek-V4-Flash-0731 — post-training upgrade of V4-Flash, benchmarks now beating GLM-5.2

TL;DR

DeepSeek announces a public-beta upgrade to the DeepSeek-V4-Flash API, keeping the exact same model architecture and size as the V4-Flash preview but shipping a substantially better post-training pass focused on agent capabilities. The tweet claims the upgraded Flash now far surpasses the earlier V4-Pro-Preview on the benchmarks in the attached image, and Thomas Neff’s Slack framing is that it beats GLM-5.2 across the board at ~3× smaller parameter count. The official V4-Flash now natively supports the OpenAI Responses API format and is adapted for Codex; V4-Pro API and app/web models are unchanged for now. No paper, model card update, or numeric benchmark table is linked in the tweet — only API docs.

Key claims

The upgrade applies only to the DeepSeek-V4-Flash API; V4-Pro API and DeepSeek’s App/Web models remain unchanged at the time of the tweet [tweet body ⚠️ Note].
DeepSeek-V4-Flash-0731 has the exact same model architecture and size as the earlier preview version — the gains come from post-training, not scaling [tweet body ⚠️ Note].
Post-training upgrade is framed as “massively upgraded Agent capabilities”, with benchmark scores in the attached image reportedly “far surpassing” V4-Pro-Preview [tweet body].
V4-Flash-0731 natively supports the Responses API format and is adapted for Codex out of the box [tweet body, linked api-docs].

Method

Tweet-only announcement; no methodology is disclosed. The relevant model-level context is that the underlying V4-Flash checkpoint is documented on the HuggingFace collection page filed as DeepSeek-V4 collection release (Flash + Pro, up to 1.6T) — 158B chat parameters against a 292B Flash-Base, released April 24, 2026. The July 31 announcement is a post-training-only refresh of that same checkpoint size, with no architecture change; the only inference-surface change is the addition of native Responses API format support and Codex compatibility. No open-weights update is announced — this refresh appears to be API-only at filing time.

Results

No numeric table is included in the tweet text; the image (not fetched) is claimed to show V4-Flash-0731 scores exceeding V4-Pro-Preview on the same benchmarks. Thomas Neff’s Slack note claims the new Flash beats GLM-5.2 across the board at ~3× smaller — independent verification would require running against SWE-Bench Pro / Terminal-Bench 2.0 / SWE-Bench Verified numbers as reported in GLM-5: from Vibe Coding to Agentic Engineering (GLM-5) and the corresponding GLM-5.2 disclosures.

Why it’s interesting

Post-training-only capability leaps at fixed parameter count and fixed architecture are the load-bearing evidence for the recipe framing in Agentic Software Engineering — that most of the recent Pareto movement in coding-agent benchmarks comes from execution-grounded synthesis + multi-stage post-training (SFT → domain experts → distillation → RL with environment feedback), not from scaling. If DeepSeek-V4-Flash-0731 replicates the Qwen3-Coder-Next-style (Qwen3-Coder-Next Technical Report) recipe at 158B active with no scale-up and lands above GLM-5.2, that is another datapoint for the concept’s central claim. Sits alongside the earlier V4 collection drop (DeepSeek-V4 collection release (Flash + Pro, up to 1.6T)) as a “same weights class, better post-training” refresh — a release pattern that Open foundation-model releases hasn’t yet tracked explicitly (most cluster contributions are new-checkpoint drops rather than post-training-only refreshes of an existing checkpoint). Complements Benchmarking Coding Agents on Databricks' Multi-Million Line Codebase‘s finding that GLM-5.2 statistically tied Claude Opus 4.8 in the top capability tier at half the cost — if V4-Flash-0731 now beats GLM-5.2 at 3× smaller size, the open-vs-closed cost/parity story shifts again.