Chinese open-weight frontier day: DeepSeek V4 Pro 0813 and Qwen3.8-2.4T land in the same 24-hour window

4 sources

On 2026-08-12 to 2026-08-13 the Chinese open-weight frontier shipped two back-to-back milestones: DeepSeek updated deepseek-v4-pro to DeepSeek-V4-Pro-0813 (called from the same model name, price unchanged; announced benchmark deltas on Terminal Bench 2.1 72.1 → 87.9, Cybergym 52.7 → 83.3, DeepSWE 12.8 → 62.7, AutomationBench 12.8 → 31.8 — reported in the announcement thread, not extracted from a DeepSeek model card); and Alibaba / Qwen released Qwen3.8-2.4T as the open-weight variant in the Qwen3.8 family, with the A95B-FP8 variant public on HuggingFace. The pattern: Chinese open-weight frontier vendors are continuing to ship aggressively, and the same-day coordination of two checkpoints turns the 8/12-8/13 window into a 'Chinese open-weight frontier accelerates' editorial frame. The takeaway for creators: the buy-side market for Chinese open-weight models is now a release-cadence market, not a vendor-decision market — pick the open-weight license that matches your team's release tolerance and let the benchmarks surface the differences.

Why now

The story is the editorial frame for 'Chinese open-weight frontier accelerates' — useful because creators waiting to pick a Chinese open-weight model now have two checkpoints in the same 24-hour window to anchor side-by-side comparisons.

Why it is worth publishing

Demo potential: a 6-8 minute explainer on what DeepSeek V4 Pro 0813 changes for agent workflows vs Qwen3.8-2.4T for open-weight serving, plus a short segment on how the two vendors are positioning differently (DeepSeek on price-stable API, Qwen on public weight license).

Evidence basis

DeepSeek API docs + HN/Algolia discussion of the 0813 checkpoint + HuggingFace model card + HN countdown-page submission

“Two Chinese open-weight frontier milestones shipped in the same 24-hour window — DeepSeek V4 Pro 0813 quietly dropped with agent-eval jumps, and Alibaba / Qwen 3.8-2.4T landed on HuggingFace as the open-weight play.”

Angle

Frame the two releases as the same editorial event — 'Chinese open-weight frontier accelerates' — and use the side-by-side as the lens for picking (DeepSeek on price-stable API, Qwen on public weight license).

Format

Long-form explainer

Demo idea

Record a 10-minute comparison: 1 min 'why both shipped today', 3 min DeepSeek V4 Pro 0813 (agent-eval deltas, price stability, who the API fits), 3 min Qwen3.8-2.4T (open-weight serving, A95B activation parameter count, who the public-weight license fits), 2 min side-by-side decision tree, 1 min 'which to pick for a Q4 build'.

Platform notes

Quote the DeepSeek benchmark deltas as 'reported in the 0813 announcement thread' rather than 'DeepSeek-reported' until the model card is on the record. Quote the Qwen3.8-2.4T activation parameter count as 'A95B' (the filename suffix) rather than as a confirmed exact parameter count until the model card is re-extracted for MoE configuration.

Usable claims

DeepSeek updated deepseek-v4-pro to DeepSeek-V4-Pro-0813 on 2026-08-13; the API calling method remains unchanged and the same model name (deepseek-v4-pro) routes to the latest version.
DeepSeek V4 Pro 0813 reports the following benchmark deltas versus the prior checkpoint: Terminal Bench 2.1 72.1 → 87.9; Cybergym 52.7 → 83.3; DeepSWE 12.8 → 62.7; AutomationBench 12.8 → 31.8. Price remains unchanged.
Alibaba's Qwen team released Qwen3.8-2.4T as the open-weight variant in the Qwen3.8 family on 2026-08-12; the A95B-FP8 variant is the public weight release on HuggingFace.

Evidence pipeline

From the news

Breakdown

Two Chinese open-weight frontier milestones shipped in the same 24-hour window — DeepSeek V4 Pro 0813 and Qwen3.8-2.4T — useful to frame as 'Chinese open-weight frontier accelerates' but easy to misread as 'one vendor caught up with the other'. The risk: the DeepSeek 0813 benchmark numbers were quoted in the announcement thread, not extracted from a DeepSeek model card, and the Qwen3.8-2.4T activation parameter count is a community inference from the filename. This explainer uses the same-day pairing as the lens to compare the two picks (DeepSeek on price-stable API, Qwen on public-weight license), while keeping the unverified benchmark numbers and inferred configuration clearly labeled.

Sources

Risks

Frame the benchmarks as 'reported in the 0813 announcement thread' or 'per the announcement'. Verify the exact numbers against a DeepSeek-published model card before quoting them as DeepSeek-reported.
Refer to the model by name and describe the release qualitatively ('public weight release for the largest Qwen3.8 variant on HuggingFace in the A95B-FP8 form'). Open the HuggingFace model card directly to extract specific numbers before stating them on the record.

Demo ideas

Side-by-side decision tree: 'DeepSeek V4 Pro 0813 vs Qwen3.8-2.4T' — pick by deployment shape (API-stable / self-hosted open-weight).
Agent-eval deep dive: what the Terminal Bench 2.1 / Cybergym / DeepSWE / AutomationBench jumps mean for a typical creator workflow (multi-step tool use, code-fix loops, web automation).
Open-weight license comparison: how Qwen3.8-2.4T's A95B-FP8 license differs from earlier Qwen licenses, and what self-hosted creators should check before download.