Skip to content

chore(models): split GLM-5.2 into its own model, deprecate GLM 5/5.1 / 将 GLM-5.2 拆分为独立模型,并弃用 GLM 5/5.1#590

Closed
functionstackx wants to merge 1 commit into
masterfrom
chore/glm52-split-deprecate-glm5
Closed

chore(models): split GLM-5.2 into its own model, deprecate GLM 5/5.1 / 将 GLM-5.2 拆分为独立模型,并弃用 GLM 5/5.1#590
functionstackx wants to merge 1 commit into
masterfrom
chore/glm52-split-deprecate-glm5

Conversation

@functionstackx

@functionstackx functionstackx commented Jul 18, 2026

Copy link
Copy Markdown
Contributor

Summary

GLM-5.2 is not a point release of GLM-5/5.1 — it is a new architecture bucket (glm_moe_dsa with IndexShare sparse attention, 78 layers; 743B core MoE / 39B active — the 753B HF metadata figure includes the bundled ~10B MTP head, same trap as DeepSeek-R1's 685B). PR #588 folded glm5.2 into the GLM-5 display grouping by mistake. This PR:

  • Splits glm5.2 into its own GLM-5.2 display model
    • DB_MODEL_TO_DISPLAY: 'glm5.2' → 'GLM-5.2' (own bucket; no re-ingest needed — DB rows were always stored under the glm5.2 key, only the display mapping changes)
    • New Model.GLM_5_2 enum + MODEL_CONFIG entry GLM5.2 743B (category default). Placed above GLM_5 because MODEL_PREFIX_MAPPING matches artifact names by substring in insertion order and glm5.2 artifact names also contain glm5 — covered by new unit tests
    • New /compare + /compare-per-dollar slug glm-5-2 (dbKeys: ['glm5.2']); the glm-5-1 slug reverts to ['glm5.1', 'glm5'] with label GLM 5/5.1. Existing glm/glm-5 aliases stay on glm-5-1, mirroring how the bare minimax alias stayed on the M2 series when M3 (different architecture) split out
    • KNOWN_MODELS, agentic sibling-nav label, and submissions display-mapping pick up GLM-5.2; /about FAQ model list derives from DB_MODEL_TO_DISPLAY automatically
  • Moves GLM 5/5.1 (Model.GLM_5) to the Deprecated dropdown group (label reverts to GLM5/5.1 744B) — no longer actively benchmarked, following the gpt-oss precedent in chore(models): move gpt-oss 120B to deprecated / 将 gpt-oss 120B 移至弃用类别 #535
  • AGENTS.md: adds the GLM-5.2 row (743B / 39B, zai-org/GLM-5.2) to the verified parameter table and a mislabel-trap bullet (not-a-point-release + 753B-includes-MTP-head)
  • Updates compare-page SEO blurbs (EN + ZH) from "GLM 5" to "GLM 5.2"

Chinese pages (/zh)

Both /zh/compare and /zh/compare-per-dollar SEO blurbs updated in the same PR. No new page/tab/post — the new compare slug pages are generated from shared templates that already have their /zh mirrors (compare-ssr-zh.ts untouched because the English narrative templates are unchanged).

Overlay-run support

N/A — this is a model-category/display-mapping change, not a chart feature. Overlay rendering resolves model labels through the same MODEL_CONFIG / DB_MODEL_TO_DISPLAY maps, so ?unofficialrun= overlays pick up the split automatically.

Testing

  • pnpm typecheck, pnpm lint, pnpm fmt — pass
  • pnpm test:unit — 3,063 tests pass, including new coverage: glm5.2 prefix-collision parsing (getModelAndSequence / getModelAndSequenceFromArtifact), GLM-5.2 own display bucket (constants + app), isModelDeprecated(GLM_5), glm-5-2 compare-slug parsing, submissions glm5.2 → GLM-5.2
  • dropdown-switching.cy.ts grouping e2e updated: GLM5/5.1 744B now asserted under Deprecated

中文说明

GLM-5.2 并非 GLM-5/5.1 的小版本更新,而是全新架构分桶(glm_moe_dsa,采用 IndexShare 稀疏注意力,78 层;核心 MoE 743B / 激活 39B —— HF 元数据显示的 753B 包含约 10B 的 MTP 头,与 DeepSeek-R1 685B 是同类陷阱)。PR #588 误将 glm5.2 并入 GLM-5 显示分组。本 PR:

  • glm5.2 拆分为独立的 GLM-5.2 显示模型DB_MODEL_TO_DISPLAY 映射为独立分桶(无需重新摄取数据,DB 行本就存储在 glm5.2 键下,仅显示映射变更);新增 Model.GLM_5_2 枚举与 MODEL_CONFIG 条目 GLM5.2 743B(默认分类),且置于 GLM_5 之上以保证产物名前缀子串匹配的正确性(已补充单元测试);新增 /compare/compare-per-dollarglm-5-2 slug,glm-5-1 slug 恢复为仅查询 ['glm5.1', 'glm5']
  • 将 GLM 5/5.1(Model.GLM_5)移入下拉框"已弃用"分组(标签恢复为 GLM5/5.1 744B),不再进行日常基准测试,沿用 chore(models): move gpt-oss 120B to deprecated / 将 gpt-oss 120B 移至弃用类别 #535 弃用 gpt-oss 的先例
  • AGENTS.md:在权威参数表中新增 GLM-5.2(743B / 39B)行,并补充参数量陷阱说明
  • 同步更新中英文 compare 页面的 SEO 描述("GLM 5" → "GLM 5.2")

测试:pnpm typecheck / pnpm lint / pnpm fmt / pnpm test:unit(3,063 个测试)全部通过;新增前缀冲突解析、独立显示分桶、弃用分类与 compare slug 的测试覆盖;已更新下拉分组 e2e。

🤖 Generated with Claude Code


Note

Medium Risk
Wide but coordinated model-taxonomy changes (dropdown categories, compare slugs, prefix parsing); regressions could mis-route benchmarks, though new tests target the main collision and mapping paths.

Overview
Treats GLM-5.2 as its own architecture and display model instead of folding glm5.2 into the GLM-5/5.1 bucket. glm5.2GLM-5.2 in shared DB/display maps; new Model.GLM_5_2 with default dropdown label GLM5.2 743B, placed above GLM_5 so artifact prefix matching does not treat glm5.2 as glm5.

GLM 5/5.1 (Model.GLM_5) moves to Deprecated (GLM5/5.1 744B); compare glm-5-1 only queries glm5 / glm5.1 and a new glm-5-2 slug targets glm5.2 alone. EN/ZH compare SEO copy highlights GLM 5.2; AGENTS.md adds parameter counts and a mislabel trap.

Tests cover prefix collision, slug parsing, deprecation, submissions URLs, and Cypress expects GLM5/5.1 744B under Deprecated.

Reviewed by Cursor Bugbot for commit 3c15315. Bugbot is set up for automated code reviews on this repo. Configure here.

GLM-5.2 is a new architecture bucket (743B core MoE / 39B active with
IndexShare DSA; the 753B HF metadata figure includes the bundled MTP
head), not a point release of GLM-5/5.1 — #588 folded it into the GLM-5
display grouping by mistake. Split glm5.2 into its own GLM-5.2 display
model with its own dropdown entry (default category) and /compare slug
(glm-5-2), and move the GLM 5/5.1 entry to the Deprecated group (no
longer actively benchmarked). Records the 753B MTP-head param trap and
the not-a-point-release trap in AGENTS.md.

中文:GLM-5.2 是全新架构分桶(核心 MoE 743B / 激活 39B,采用 IndexShare
稀疏注意力;HF 元数据显示 753B 是因为包含 MTP 头),并非 GLM-5/5.1 的
小版本更新 — #588 误将其并入 GLM-5 显示分组。本次将 glm5.2 拆分为独立
的 GLM-5.2 显示模型(默认分类,含独立下拉项与 /compare slug glm-5-2),
并将 GLM 5/5.1 条目移入"已弃用"分组(不再进行日常基准测试);同时在
AGENTS.md 中记录 753B(含 MTP 头)参数量陷阱与"非小版本更新"陷阱。

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
@functionstackx
functionstackx requested a review from adibarra as a code owner July 18, 2026 20:32
@vercel

vercel Bot commented Jul 18, 2026

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated (UTC)
inferencemax-app Ready Ready Preview, Comment Jul 18, 2026 8:33pm

Request Review

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant