Add Qwen3.8-Flash-Next (FP8 + NVFP4) / 新增 Qwen3.8-Flash-Next 模型（FP8 与 NVFP4）

[functionstackx]

Summary

Registers Qwen3.8-Flash-Next — the Qwen4-architecture preview — as its own model, wired end to end: DB bucket, normalizer paths, dashboard selector, compare pages, rankings, and run pages.

DB key / infmax_model_prefix	qwen3.8next
Display name	Qwen3.8-Flash-Next
Dashboard label	Qwen3.8 Flash Next 176B
Compare slug	qwen-3-8-flash-next
Category	experimental (see below)
Release date	2026-08-26

Both serving paths resolve, and both were verified against real sweeps

Hopper has no NVFP4 path, so H200 serves the FP8 checkpoint while Blackwell serves NVFP4. I checked each against an executed run rather than inferring either from a PR description:

HF path	→ key	Verified in
Qwen/Qwen3.8-Flash-Next-FP8	qwen3.8next	run 33038487711 (H200, #2753)
RadixArk/Qwen3.8-Flash-Next-NVFP4	qwen3.8next	run 33039186365 (B300, #2752)

Resolution exercised directly through resolveModelKey, covering the row shapes the ingest actually sees:

{ model: 'Qwen/Qwen3.8-Flash-Next-FP8' }         -> qwen3.8next
{ model: 'RadixArk/Qwen3.8-Flash-Next-NVFP4' }   -> qwen3.8next
{ infmax_model_prefix: 'qwen3.8next' }           -> qwen3.8next
{ model: 'Qwen/Qwen3.5-397B-A17B-FP8' }          -> qwen3.5      (unchanged)

Both runs report infmax_model_prefix: qwen3.8next, so no PREFIX_ALIASES entry is needed. The fp4 precision the B300 job reports is already in KNOWN_PRECISIONS.

The parameter count is 176B, not 125B

The model card leads with "125B with 6B activated", but that covers the main model only. The 51B n-gram embedding table brings the total to 176B, and a separate 4B MTP head sits outside both figures.

This is exactly the class of mistake the AGENTS.md parameter table exists to prevent, so the row and a mislabel-trap entry are added there alongside the GLM-5 and MiniMax entries.

Architecture, for the record: GatedDeltaNet + Qwen Sparse Attention hybrid, 512 experts (10 routed + 1 shared), license qwen-community-1.0.

Why experimental rather than default

Every other current model is default, so this is a deliberate exception worth a second opinion.

The only data today is the day-zero agentic arm. default puts a model into DEFAULT_MODELS, which is what builds the /overview matrix — and I confirmed against the overview fixture that a default registration emits a Qwen3.8-Flash-Next/single_turn_8k1k row with no data behind it. MODEL_OPTIONS excludes only hidden, so experimental keeps the model fully selectable in the dashboard, compare pages, rankings, and run pages while keeping that empty row out of the curated matrix.

Promoting it later is a one-word change in MODEL_CONFIG, and the comment there says so.

On the release date

The model card states no publication date. Qwen announced the open-sourcing for 23:00 Beijing time on 2026-08-26, which is also the day the lmsysorg/sglang:qwen38flashnext bring-up tag was published. Some coverage puts the Hugging Face repo live on 08-24; either way it precedes the first sweep on 08-27, which is the invariant model-architectures.test.ts pins. The reasoning is recorded in a comment next to the entry, matching the style of its neighbours.

Three pinned-count guards moved

These are guards that exist to force a decision when a model is added, not incidental churn:

• compare-slug.test.ts — seoName coverage map gains the new slug
• rankings.test.ts — 22 → 24 (2 kinds × 12 models)
• run-pages.test.ts — 99 → 108 (12 models × 9 chips)

The api-route-catalog digest for constants/src/models.ts is refreshed. Per the rule in AGENTS.md I checked the documentation first rather than bumping the hash blind: api-documentation.ts derives its model enum from DB_MODEL_TO_DISPLAY, so the new key appears automatically and no reference copy, example, or OpenAPI schema needed editing.

Not done here

• model-architectures.ts — optional per the checklist and skipped deliberately. The entries carry verified config.json values, and this model's shape (n-gram embedding table, separate MTP head, QSA) is unusual enough that I would rather add it from the config than approximate it. Omitted models simply render no diagram.
• The /compare SEO blurb lists a five-model sample and is already 158 characters; adding a sixth name would push the meta description past the useful length. Left alone.

Deploy order matters. Per docs/adding-entities.md, the packages/constants and packages/db changes must be merged and deployed before the first ingest containing qwen3.8next rows, or those rows are silently skipped by the normalizer and need a full re-ingest to recover.

Verification

• bun run test:unit — 4,419 app tests pass (246 files), all workspaces green
• bun run typecheck clean — the Record<Model, …> configs are the safety net that catches a half-registered enum member
• bun run build — zero errors; the new model appears in INFERENCE_MODEL_SLUGS (12 entries) and generates 9 run pages
• resolveModelKey exercised directly on both HF paths and the artifact prefix

中文说明

新增 Qwen3.8-Flash-Next（Qwen4 架构预览）作为独立模型，并完成端到端接入：数据库分桶、normalizer 路径、仪表板选择器、对比页、排名页与运行页。DB key 为 qwen3.8next，展示名 Qwen3.8-Flash-Next，标签 Qwen3.8 Flash Next 176B，对比 slug 为 qwen-3-8-flash-next，分类 experimental，发布日期 2026-08-26。

两条服务路径均已通过真实 sweep 验证

Hopper 无 NVFP4 路径，因此 H200 使用 FP8 权重、Blackwell 使用 NVFP4。两者均对照已执行的运行核实，而非依据 PR 描述推断：Qwen/Qwen3.8-Flash-Next-FP8 来自 run 33038487711（H200，#2753），RadixArk/Qwen3.8-Flash-Next-NVFP4 来自 run 33039186365（B300，#2752）。两次运行的 infmax_model_prefix 均为 qwen3.8next，因此无需 PREFIX_ALIASES；B300 任务上报的 fp4 精度已在 KNOWN_PRECISIONS 中。已直接调用 resolveModelKey 验证各种 row 形态，且 Qwen3.5 的解析结果保持不变。

参数量是 176B 而非 125B

model card 首行写的是「125B with 6B activated」，但那只是主模型；加上 51B n-gram embedding 后总量为 176B，另有 4B MTP 头不计入其中。这正是 AGENTS.md 参数表要防范的错误类型，因此已在该表及易错清单中补充相应条目。架构为 GatedDeltaNet + Qwen Sparse Attention 混合结构，512 个专家（10 路由 + 1 共享），许可证 qwen-community-1.0。

为何使用 experimental 而非 default

其他现有模型均为 default，因此这是一个需要评审确认的有意例外。目前仅有首发 agentic 数据；default 会把模型纳入 DEFAULT_MODELS，而 /overview 矩阵正是由它构建——我已对照 overview fixture 确认，设为 default 会生成一条没有数据支撑的 Qwen3.8-Flash-Next/single_turn_8k1k 行。MODEL_OPTIONS 仅排除 hidden，因此 experimental 仍可在仪表板、对比页、排名页和运行页中正常选择。后续提升为 default 只需改动一个词。

关于发布日期

model card 未标注发布日期。Qwen 宣布于北京时间 2026-08-26 23:00 开源，同日发布了 lmsysorg/sglang:qwen38flashnext 适配镜像标签；部分报道称 Hugging Face 仓库在 08-24 上线。无论采用哪个，都早于 08-27 的首次 sweep，满足 model-architectures.test.ts 所固定的不变式。相关推理已按邻近条目的风格写入注释。

三处计数守卫按预期变化

compare-slug.test.ts 的 seoName 覆盖表新增该 slug；rankings.test.ts 由 22 增至 24（2 种类型 × 12 个模型）；run-pages.test.ts 由 99 增至 108（12 个模型 × 9 款芯片）。constants/src/models.ts 的 api-route-catalog 摘要哈希已更新——按 AGENTS.md 规定先核查了文档：api-documentation.ts 的模型枚举派生自 DB_MODEL_TO_DISPLAY，新 key 会自动出现，无需修改参考文档、示例或 OpenAPI schema。

本次未做

model-architectures.ts 属可选项，本次有意跳过：该文件条目均使用经核实的 config.json 数值，而该模型结构较特殊（n-gram embedding 表、独立 MTP 头、QSA），宁可日后依据 config 补充，也不做近似；缺失条目仅表现为不渲染架构图。/compare 的 SEO 简介目前列举五个模型、已达 158 字符，再加一个会超出 meta description 的合理长度，故未改动。

部署顺序很重要。 按 docs/adding-entities.md，packages/constants 与 packages/db 的改动必须在首次包含 qwen3.8next 数据的 ingest 之前合并并部署，否则这些行会被 normalizer 静默跳过，需要完整重新 ingest 才能恢复。

验证

• bun run test:unit — app 端 4,419 项测试通过（246 个文件），各 workspace 全部通过
• bun run typecheck 通过——Record<Model, …> 类型是捕获枚举成员未完整注册的安全网
• bun run build — 零报错；新模型已出现在 INFERENCE_MODEL_SLUGS（共 12 项）并生成 9 个运行页
• 已直接对两条 HF 路径与 artifact prefix 调用 resolveModelKey 验证

🤖 Generated with Claude Code

---

Note

Medium Risk
Touches the ETL model-resolution path and shared model registries; incorrect deploy order or mapping would drop or mis-attribute qwen3.8next ingest rows until re-ingest.

Overview
Registers Qwen3.8-Flash-Next as its own model (qwen3.8next / display Qwen3.8-Flash-Next) across ingest, shared constants, the dashboard Model enum, compare URLs (qwen-3-8-flash-next), and SSR allowlists. ETL maps both Qwen/Qwen3.8-Flash-Next-FP8 and RadixArk/Qwen3.8-Flash-Next-NVFP4 into the same DB bucket.

The dashboard lists it as Qwen3.8 Flash Next 176B with category experimental so it stays selectable but is excluded from the /overview default matrix until more sweep coverage exists. AGENTS.md documents the 176B total (not 125B) and adds a mislabel trap.

Registry tests bump expected ranking pages (22→24) and run pages (99→108); api-route-catalog refreshes the digest for constants/src/models.ts after the new display mapping and release date (2026-08-26).

^{Reviewed by Cursor Bugbot for commit 91a1dd7. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
inferencemax-app	Ready	Preview	Aug 27, 2026 5:45am