[Klaud Cold] docs: add MODELS.md model & scenario support matrix / 新增 MODELS.md 模型与场景支持矩阵 - #2343
Conversation
Add MODELS.md (+ MODELS_zh.md) documenting every benchmarked model: date added, active scenarios, and deprecated scenarios. Single-turn 1k1k and 1k8k are marked deprecated for all models; retired models (llama70b, gptoss, minimaxm2.5, glm5, glm5.1) list all their scenarios as deprecated. Register the new bilingual doc pair in AGENTS.md. 中文：新增 MODELS.md 与 MODELS_zh.md，记录所有被基准测试的模型：加入日期、启用场景与已弃用场景。单轮 1k1k 与 1k8k 场景对所有模型均标记为弃用；已退役模型（llama70b、gptoss、minimaxm2.5、glm5、glm5.1）的全部场景均标记为弃用。同时在 AGENTS.md 中登记该双语文档对。 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Note in the scenario table that single-turn 1k1k and 1k8k were deprecated to save GPU cluster time for higher-priority real-world agentic-coding benchmarks and new frontier models. 中文：在场景表中注明单轮 1k1k 与 1k8k 弃用的原因——将 GPU 集群时间留给优先级更高的真实场景智能体编码基准测试与新的前沿模型。 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…[skip-sweep] Rename the support-matrix Model column to 'Model architecture class' and collapse GLM-5 and GLM-5.1 into a single row (both prefixes, both add dates, shared retirement). 中文：将支持矩阵中的「模型」列重命名为「模型架构类别」，并将 GLM-5 与 GLM-5.1 合并为一行（含两个前缀、各自加入日期与共同退役信息）。 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
中文：将模型支持矩阵按加入日期从新到旧排序。 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…p-sweep] GLM-5.2 postdates both deprecations, and dsv4 / minimaxm3 / glm5.1 were added after 1k8k was removed in #911 — a scenario a model never ran is not listed as deprecated for it. Also drop the dsr1 disabled-agentic footnote. 中文：仅对实际运行过 1k1k/1k8k 的模型将其列为已弃用场景。GLM-5.2 晚于两个场景的弃用；dsv4、minimaxm3、glm5.1 在 #911 移除 1k8k 之后才加入——模型从未运行过的场景不再标记为弃用。同时删除 dsr1 智能体编码配置暂停用的脚注。 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… [skip-sweep] MiniMax-M2.5 never ran the agentic-coding scenario, so it is not listed as deprecated for it. GLM-5.2's deprecated-scenarios cell is now simply blank instead of carrying an explanatory dash note. 中文：MiniMax-M2.5 从未运行过智能体编码场景，故不再将其列为已弃用；GLM-5.2 的已弃用场景单元格改为留空，不再附加说明。 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
| | Single-turn 1k1k | 1024 / 1024 | **Deprecated for all models** since 2026-07-17 ([#2263](https://github.com/SemiAnalysisAI/InferenceX/pull/2263)). Archived configs live in [`configs/deprecated/`](configs/deprecated/). | | ||
| | Single-turn 1k8k | 1024 / 8192 | **Deprecated for all models** since 2026-03-27 ([#911](https://github.com/SemiAnalysisAI/InferenceX/pull/911)). Configs were removed, not archived. | | ||
| | Single-turn 8k1k | 8192 / 1024 | Active — the primary fixed-sequence-length scenario. | | ||
| | Agentic coding | trace replay | Active — trace-replay agentic-coding benchmark (see [`benchmarks/agentic/`](benchmarks/agentic/)). | |
There was a problem hiding this comment.
🟡 MODELS.md (line 14) and MODELS_zh.md (line 14) link the Agentic coding scenario to benchmarks/agentic/, but that directory doesn't exist in the repo — the link 404s on GitHub. The actual agentic benchmark entrypoints are benchmarks/single_node/agentic/ and benchmarks/multi_node/agentic/ per AGENTS.md's Directory Structure section; both docs should point there instead.
Extended reasoning...
Both new docs added by this PR — MODELS.md and its Chinese counterpart MODELS_zh.md — contain a broken relative link. Line 14 of each file (the "Agentic coding" row in the Scenarios table) reads:
| Agentic coding | trace replay | Active — trace-replay agentic-coding benchmark (see [`benchmarks/agentic/`](benchmarks/agentic/)). |
benchmarks/agentic/ does not exist anywhere in the repository. I verified this directly against the working tree:
$ find benchmarks -maxdepth 2 -type d -iname '*agentic*'
benchmarks/multi_node/agentic
benchmarks/single_node/agentic
There is no top-level benchmarks/agentic/ — only the nested single_node/agentic/ and multi_node/agentic/ directories exist, matching what AGENTS.md's own Directory Structure section documents ("single_node/ and multi_node/ entrypoints"). Since GitHub renders relative Markdown links against the file's path in the repo tree, [benchmarks/agentic/](benchmarks/agentic/) resolves to a real repo path that 404s when clicked.
Why this slipped through: the rest of the doc's relative links are correct — e.g. [configs/deprecated/](configs/deprecated/) on the two lines directly above resolves to a real directory, so this looks like a simple oversight where the author collapsed the two real agentic subdirectories into a single simplified (but nonexistent) path when writing the doc.
Impact: this is a new, docs-only file with no functional/runtime effect — nothing in CI, benchmarking, or config generation reads this link. A reader who clicks through from the published doc to find the agentic benchmark code hits a 404 instead of the actual scripts, which is a minor rough edge for a reference doc whose whole purpose is to be a navigable index.
Fix: point the link at both real entrypoints, e.g. (see [benchmarks/single_node/agentic/](benchmarks/single_node/agentic/) and [benchmarks/multi_node/agentic/](benchmarks/multi_node/agentic/)), applied identically to MODELS.md line 14 and MODELS_zh.md line 14 to keep the bilingual docs in sync per AGENTS.md's mirroring requirement.
…sweep] Rename the architecture-class labels to cover all variants benchmarked under each prefix: MiniMax-M2.5/2.7 (minimaxm2.5) and Kimi-K2.5/2.6/2.7-Code (kimik2.5). 中文：更新架构类别名称以涵盖各前缀下基准测试过的全部变体：MiniMax-M2.5/2.7（minimaxm2.5）与 Kimi-K2.5/2.6/2.7-Code（kimik2.5）。 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Kimi-K3 (kimik3, added 2026-07-27) and Qwen3.8 2.4T (qwen3.8, date TBD), both agentic-coding only. 中文：在模型矩阵中新增 Kimi-K3（kimik3，2026-07-27 加入）与 Qwen3.8 2.4T（qwen3.8，日期待定），两者均仅运行智能体编码场景。 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Describe the agentic-coding scenario as long-context, multi-turn realistic-traffic trace replay with sub agents, and note that new models will likely be onboarded with agentic coding only. 中文：将智能体编码场景描述为长上下文、多轮真实流量的轨迹回放（含子智能体），并注明今后新模型预计仅以智能体编码场景接入。 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Reorder the scenario table: agentic coding, single-turn 8k1k, then the deprecated single-turn 1k1k and 1k8k. 中文：按优先级重排场景表：智能体编码、单轮 8k1k，随后为已弃用的单轮 1k1k 与 1k8k。 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
1 participant
Summary
Adds
MODELS.md(with its bilingual counterpartMODELS_zh.md) documenting every model InferenceX has benchmarked, as a single support-matrix table with:Retired models are included with all their scenarios marked deprecated:
llama70b(configs deleted outright, pre-datingconfigs/deprecated/),gptoss(#2101),minimaxm2.5(#1874),glm5/glm5.1(#2276).Also registers the new
MODELS.md/MODELS_zh.mdpair in the bilingual-docs list inAGENTS.md.Docs-only change — no configs or
perf-changelog.yamltouched, head commit carries[skip-sweep].中文说明
新增
MODELS.md（及其中文版MODELS_zh.md），以支持矩阵表格的形式记录 InferenceX 基准测试覆盖过的所有模型，包括：已退役模型一并收录，其全部场景均标记为弃用：
llama70b（配置被直接删除，早于configs/deprecated/归档机制）、gptoss（#2101）、minimaxm2.5（#1874）、glm5/glm5.1（#2276）。同时在
AGENTS.md的双语文档列表中登记MODELS.md/MODELS_zh.md文档对。仅文档变更 — 未改动任何配置或
perf-changelog.yaml，head commit 带有[skip-sweep]标记。🤖 Generated with Claude Code