[Klaud Cold] docs: deprecate DeepSeek-V4-Pro Single-turn 8k1k on 2026-09-08, keep AgentX / 弃用 DeepSeek-V4-Pro 单轮 8k1k（2026-09-08），保留 AgentX

[functionstackx]

Summary / 摘要

Announces Tuesday, September 8, 2026 (two weeks out) as the last day for the Single-turn 8k1k scenario on DeepSeek-V4-Pro 1.6T (dsv4). Agentic coding / AgentX stays active for dsv4, including the MTP and DSpark arms, so the model is not retired — agentic coding simply becomes its only scenario.

宣布 2026 年 9 月 8 日（星期二）（即两周后）为 DeepSeek-V4-Pro 1.6T（dsv4）单轮 8k1k 场景的最后运行日。智能体编码 / AgentX 对 dsv4 继续启用，含 MTP 与 DSpark 分支，因此该模型不会退役，智能体编码将成为其唯一场景。

Changes / 改动

• MODELS.md: new ### Tuesday, September 8, 2026 section in the Deprecation Notice with the deprecated/remains table, rationale, and a Status: not yet enacted note; dsv4 row of the model support matrix now flags the pending last day.

• MODELS_zh.md: synchronized Chinese mirror.

• MODELS.md：在弃用公告中新增 ### 2026 年 9 月 8 日（星期二） 小节，包含弃用/保留对照表、原因说明，以及状态：尚未执行的说明；模型支持矩阵的 dsv4 行标注该最后运行日。

• MODELS_zh.md：同步中文镜像。

Rationale / 原因

dsv4 has the largest single-turn footprint in the repo: 45 active config keys use 8k1k (32 in configs/nvidia-master.yaml, 13 in configs/amd-master.yaml) across H200, B200, B300, GB200, GB300, MI300X, MI325X, MI355X and vLLM, SGLang, TensorRT-LLM, ATOM, Dynamo, llm-d. Its 19 agentic config keys are the part that feeds the published North Star Pareto frontier. Retiring the fixed-sequence-length arm frees cluster hours for AgentX and new frontier models without reducing what we publish for this model.

dsv4 是仓库中单轮场景占用最大的模型：45 个启用配置项使用 8k1k（configs/nvidia-master.yaml 32 个，configs/amd-master.yaml 13 个），覆盖 H200、B200、B300、GB200、GB300、MI300X、MI325X、MI355X 及 vLLM、SGLang、TensorRT-LLM、ATOM、Dynamo、llm-d。其19 个智能体编码配置项才是支撑已发布北极星帕累托前沿的部分。下线固定序列长度分支可为 AgentX 与新前沿模型腾出集群机时，同时不减少该模型对外发布的内容。

Scope / 范围

Docs only. No config, recipe, or script changes — nothing is removed in this PR; enactment (archiving to configs/deprecated/) is a follow-up after the last day. No perf-changelog.yaml entry required.

仅文档改动。不涉及配置、配方或脚本；本 PR 不移除任何内容，实际执行（归档至 configs/deprecated/）将在最后运行日之后由后续 PR 完成。无需新增 perf-changelog.yaml 条目。

🤖 Generated with Claude Code

---

Note

Low Risk
Documentation-only policy announcement; no configs, recipes, or benchmark scripts are modified.

Overview
Documents Tuesday, September 8, 2026 as the last day for Single-turn 8k1k on DeepSeek-V4-Pro 1.6T (dsv4). Agentic coding / AgentX stays active (including MTP and DSpark); the model is not retired—only the fixed-sequence-length arm is scheduled to go away.

Adds a new Deprecation Notice subsection in MODELS.md and MODELS_zh.md with a deprecated/remains table, rationale (45 active 8k1k keys vs. 19 agentic keys on the North Star frontier), and Status: not yet enacted with follow-up steps aligned to #2493/#2527 (archive master configs and scripts; keep SPEED-Bench AL scripts). Updates the dsv4 row in the English model support matrix to flag the pending last day.

No runtime changes in this PR—configs and benchmarks still run until a later enactment PR after the cutoff.

^{Reviewed by Cursor Bugbot for commit 1232e2a. Bugbot is set up for automated code reviews on this repo. Configure here.}

[claude]

Additional findings (outside the current diff — GitHub can't attach inline comments there):

• 🟡 MODELS_zh.md — MODELS_zh.md's model-support-matrix row for dsv4 still reads "单轮 8k1k、智能体编码" without the new "(last day 2026-09-08, see Deprecation Notice)" annotation that MODELS.md's corresponding row (line ~156) gained in this diff, so the English and Chinese docs now disagree on dsv4's 8k1k deprecation status.

  Extended reasoning...

  A Chinese-reading user consulting MODELS_zh.md's support matrix sees no indication that Single-turn 8k1k for dsv4 is scheduled for deprecation on 2026-09-08, missing the notice that MODELS.md readers get, and may keep relying on the 8k1k scenario past its announced last day without warning.

  Verification: nit / documentation-parity regression introduced by this change. In the diff, MODELS.md's dsv4 matrix row gained the annotation: "Single-turn 8k1k (last day 2026-09-08, see Deprecation Notice), Agentic coding...". The MODELS_zh.md diff, however, only adds the new "2026 年 9 月 8 日" section and does NOT touch the matrix row. MODELS_zh.md:156 still reads "单轮 8k1k、智能体编码（非 MT

[claude]

🔴 New Sept 8 2026 deprecation entry claims dsv4's agentic coding 'stays active... including its MTP and DSpark arms', but no dsv4 config key in configs/nvidia-master.yaml or configs/amd-master.yaml uses DSpark (grep for 'dspark' in either file returns zero dsv4 hits, unlike kimik3 which has multiple active -agentic-dspark- keys). The pre-existing spec-decode A/B table at MODELS.md lines 27-33 also lists only 'MTP' as dsv4's current agentic spec-decode arm, not DSpark; DSpark for dsv4 is only an 'agreed PoR' (line 123) and a SPEED-Bench golden-AL reference (golden_al_distribution/dsv4-pro-0813-dspark.yaml), not a running master-config arm.

Extended reasoning...

A reader trusts this new notice and expects published dsv4 DSpark agentic results after Sept 8 2026 deprecation of Single-turn 8k1k; no such active arm exists in either master config, so the claim is currently false and will stay false unless a real dsv4-agentic-dspark key is added. Same incorrect claim is duplicated in MODELS_zh.md lines 47 and 51.

Verification: normal — the new Sept 8 2026 entry introduces a false, self-contradictory claim not present on the base branch. MODELS.md:47 and the table row at MODELS.md:51 both state dsv4's agentic coding stays active "including the MTP and DSpark arms." No dsv4 config key uses DSpark: a grep for dsv4.*dspark/dspark.*dsv4 across configs/nvidia-master.yaml and configs/amd-master.yaml returns zero matches,