Ship the Qwen3.8-Flash-Next model page, default and AgentX-only / 新增 Qwen3.8-Flash-Next 模型页面，并改为 default 与仅 AgentX 场景

[functionstackx]

Summary

Three changes that finish the Qwen3.8-Flash-Next addition started in #887.

1. /model/qwen-3-8-flash-next was a 404 — fixed

getModelPageSlugs() filters the compare slugs down to those that have a content/models/<slug>.mdx, and the model had none, so the route never generated. Adds the page, following the shape of the existing model deep-dives (Overview → Architecture → eval scores → benchmark explanations → caveats).

Verified: the page prerenders to .next/server/app/model/qwen-3-8-flash-next.html with all six sections present.

2. A Qwen4 architecture section, as the centrepiece

You asked for depth here, and it is the right emphasis: Qwen positions this checkpoint as "architectural innovation for sustainable AGI progress, departing from the scaling-focused approach of simply increasing parameters and context windows." The architecture is the reason to read about the model, so it gets its own section rather than a bullet list:

• N-gram embedding as a second parameter-scaling axis — 20,000,000 bigram/trigram embeddings at layer 2, contributing 51B parameters, which Qwen calls "a unique axis for parameter scaling that requires less computation" than MoE. The page draws out why that matters operationally: MoE buys capacity by activating more experts, paying memory bandwidth and all-to-all traffic; an n-gram table buys capacity with a lookup. On a part where HBM capacity binds before FLOPS, those are different trades — and it means "6B active" understates the memory footprint badly.
• Qwen Sparse Attention — selects at "the micro-block level" under a fixed budget of "512 blocks or 2048 tokens", via an MQA indexer with 4 query heads and 1 shared key head. The page explains the consequence: once the budget binds, attention cost stops tracking context length, so long-context interactivity stops degrading the way dense attention does.
• Gated residual — "element-wise, data-dependent read gate and per-branch scalar write gate" over 4 branches at bottleneck rank 320, with a note on why the read/write asymmetry is the design.
• The inherited pieces — the 3:1 Gated DeltaNet to full-attention interleave carried from Qwen3.5, and the 1-layer multi-step MTP head, plus why AgentX applies a separately measured acceptance length to it rather than letting synthetic replay content decide.
• Training recipe — Muon and AdamW on specific weight categories, no batch-size warmup.

Every figure is cited to the model card. The vendor eval table carries the Nemotron 3 Ultra and % better columns per the convention merged in #876 — GPQA Diamond 91.7 (+5.4%), LiveCodeBench v6 91.9 (+3.3%), SWE-bench Multilingual 81.0 (+19.6%), IFBench 81.3 (−0.5%), with n.a. on the twelve benchmarks NVIDIA does not report.

3. default instead of experimental, and AgentX-only

Per your call, the model is promoted to default so it appears in the /overview matrix from day zero.

default on its own would have given it a single_turn_8k1k row, because an uncurated model falls back to the single data-derived scenario — and this model is benchmarked on agentic traces. So it also joins Kimi K3 and GLM 5.2 in OVERVIEW_MODEL_SCENARIOS as ['agentx'], and in the fallback that picks a scenario for a model with no matching rows yet. It now contributes exactly one matrix row, on AgentX, rather than an empty fixed-sequence row.

Notes for reviewers

• The 125B trap is called out in the page itself, not just in code comments. The card leads with "125B with 6B activated", which is the main model only; 176B is the checkpoint. The caveats section says so explicitly.
• The Qwen blog is a client-rendered app and returns no extractable content, so every architectural claim is sourced to the Hugging Face model cards instead, plus TechNode for the release timing.
• Vendor-reported throughout. Qwen's own card compares against Qwen3.8-27B, Qwen3.7-Plus and DeepSeek-V4-Flash; those columns are dropped in favour of the cross-page Nemotron reference, and the caveats say the scores are self-reported under Qwen's harness.
• The api-route-catalog digest for overview-data.ts is refreshed. Per AGENTS.md I checked first: the change touches only the curated model-to-scenario map, no BFF parameter or OverviewPageData field moved, and the API reference enumerates neither.

Verification

• bun run test:unit — 4,419 app tests pass (246 files), all workspaces green
• bun run build — zero errors; /model/qwen-3-8-flash-next and its OG image both prerender
• Rendered HTML checked for the new sections and the computed deltas
• lint, fmt, typecheck clean, also via the pre-commit hook

中文说明

本 PR 完成 #887 开始的 Qwen3.8-Flash-Next 接入，共三处改动。

1. 修复 /model/qwen-3-8-flash-next 的 404

getModelPageSlugs() 只保留存在 content/models/<slug>.mdx 的 slug，而该模型此前没有内容文件，因此路由从未生成。本次新增该页面，沿用既有模型深度页的结构（概览 → 架构 → 厂商评测 → 基准说明 → 注意事项）。已验证页面成功预渲染为 .next/server/app/model/qwen-3-8-flash-next.html，六个章节齐全。

2. 以 Qwen4 架构一节为核心

这一强调是恰当的：Qwen 将该 checkpoint 定位为「面向可持续 AGI 进展的架构创新，而非单纯增加参数与上下文窗口的规模路线」。架构正是这个模型值得一读的原因，因此单独成节而非罗列要点：

• n-gram embedding 作为第二条参数扩展轴 —— 第 2 层的 2000 万条 bigram/trigram、贡献 51B 参数，Qwen 称其为「一条计算量更低的参数扩展轴」。页面进一步说明其工程意义：MoE 通过激活更多专家换取容量，代价是显存带宽与 all-to-all 流量；n-gram 表则以查表换容量。在 HBM 容量先于算力成为瓶颈的硬件上，两者是完全不同的取舍，也意味着「6B 激活」会严重低估显存占用。
• Qwen Sparse Attention —— 在「512 blocks 或 2048 tokens」的固定预算下按微块选择，索引器为 4 query head、1 共享 key head 的 MQA。页面点明其后果：预算一旦生效，注意力开销不再随上下文长度增长，长上下文交互性也不会像稠密注意力那样劣化。
• Gated residual —— 4 分支、瓶颈秩 320，采用元素级数据相关读门与每分支标量写门，并解释这种读写不对称正是设计意图。
• 沿用的部分 —— 承自 Qwen3.5 的 3:1 Gated DeltaNet 混合结构，以及单层多步 MTP 头，并说明 AgentX 为何对其套用单独测量的接受长度。
• 训练配方 —— Muon 与 AdamW 按权重类别分别应用，取消 batch size 预热。

所有数据均引用 model card。厂商评测表按 #876 的约定附带 Nemotron 3 Ultra 与 % better 两列：GPQA Diamond 91.7（+5.4%）、LiveCodeBench v6 91.9（+3.3%）、SWE-bench Multilingual 81.0（+19.6%）、IFBench 81.3（−0.5%），NVIDIA 未公布的十二项标为 n.a.。

3. 改为 default，且仅 AgentX 场景

按你的决定，模型提升为 default，自首发起即出现在 /overview 矩阵中。但仅改为 default 会让它落入 single_turn_8k1k 行，因为未经策展的模型会回退到唯一由数据推导的场景，而该模型是基于 agentic trace 测试的。因此它同时与 Kimi K3、GLM 5.2 一并在 OVERVIEW_MODEL_SCENARIOS 中登记为 ['agentx']，并同步调整了尚无匹配数据时选择场景的回退分支，最终只贡献一行 AgentX 矩阵行。

评审提示

• 125B 这个坑已写入页面本身，而不只是代码注释：card 首行的「125B with 6B activated」仅指主模型，checkpoint 实为 176B，注意事项中已明确说明。
• Qwen 官方博客是客户端渲染应用，无法抓取内容，因此所有架构结论均引用 Hugging Face model card，发布时间引用 TechNode。
• 全部为厂商自报数据。 Qwen 自家 card 的对比列为 Qwen3.8-27B、Qwen3.7-Plus 与 DeepSeek-V4-Flash，本页改用跨页面统一的 Nemotron 参照，并在注意事项中说明分数由厂商在自有 harness 下自报。
• overview-data.ts 的 api-route-catalog 摘要哈希已更新；按 AGENTS.md 先行核查：本次仅改动模型到场景的策展映射，未涉及任何 BFF 参数或 OverviewPageData 字段，API 参考也未枚举这些内容。

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 build — 零报错；/model/qwen-3-8-flash-next 及其 OG 图均成功预渲染
• 已检查渲染后的 HTML，确认新增章节与计算出的差距数值
• lint、fmt、typecheck 均通过（pre-commit hook 亦已执行）

🤖 Generated with Claude Code

---

Note

Low Risk
Documentation and curated overview scenario mapping only; no API contract or benchmark pipeline changes beyond showing one additional default-matrix row.

Overview
Adds content/models/qwen-3-8-flash-next.mdx so /model/qwen-3-8-flash-next can prerender (it was missing from getModelPageSlugs()’s MDX filter). The page follows the usual deep-dive shape: overview, Qwen4 architecture (n-gram table, QSA, gated residual, hybrid/MTP), vendor eval table with Nemotron 3 Ultra / % better columns, benchmark notes, and caveats (176B vs 125B headline, chip precision split, MTP acceptance length).

Overview integration: Qwen3.8-Flash-Next moves from experimental to default in data-mappings.ts so it appears in the default matrix. It is curated as AgentX-only in OVERVIEW_MODEL_SCENARIOS and the scenario fallback (with Kimi K3 and GLM 5.2), avoiding an empty single_turn_8k1k row. Unit tests and Cypress expect 9 matrix rows; api-route-catalog digest for overview-data.ts is refreshed.

^{Reviewed by Cursor Bugbot for commit b25c8be. 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 6:33am