Add DeepSeek-V4.1-Flash / 新增 DeepSeek-V4.1-Flash

[Oseltamivir]

Summary

Registers DeepSeek-V4.1-Flash as a dashboard model, sourced from the GB300 AgentX sweep in SemiAnalysisAI/InferenceX#2961 — DB key dsv41flash, HF path deepseek-ai/DeepSeek-V4.1-Flash, FP4 / vLLM / TP4 / MTP on gb300-nv, concurrencies 1–128.

Preview of the source run: https://inferencex.semianalysis.com/inference?unofficialRun=34463351592

Changes

File	Change
packages/constants/src/models.ts	dsv41flash → 'DeepSeek-V4.1-Flash' + release date 2026-09-10
packages/db/src/etl/normalizers.ts	'deepseek-ai/DeepSeek-V4.1-Flash' → 'dsv41flash'
packages/app/src/lib/data-mappings.ts	Model enum member + MODEL_CONFIG entry
packages/app/src/lib/compare-slug.ts	deepseek-v41-flash slug
packages/app/src/lib/compare-ssr.ts	KNOWN_MODELS entry so ?g_model= validates
packages/app/src/lib/overview-data.ts	curated AgentX-only in the overview matrix
packages/app/src/lib/api-route-catalog.ts	refreshed two shared-source digests
4 test files	updated pinned expectations (see below)

Decisions worth reviewing

• Own DB bucket, not a dsv4 point release. V4.1-Flash is a Causal Encoder-Decoder architecture with CSA2 sparse attention — structurally different from V4-Pro, so grouping them under one display name would misrepresent both. Same treatment as kimik3 vs the K2 series and minimaxm3 vs the M2 series.
• Label reads 552B, not 748B. The HF card gives a 552B backbone (8B active in prefill, 16B in decode) plus a 196B Engram conditional-memory table. Engram is sparsely accessed via token lookup rather than being resident per-token compute, so it is excluded — consistent with excluding separate MTP heads elsewhere. Flagging this explicitly because the Qwen3.8-Flash-Next precedent went the other way (its 51B n-gram embedding table is folded into the 176B total). If you read Engram as the same kind of thing, the label should become 748B.
• Release date 2026-09-10. Hugging Face repo initial commit at 02:17Z today; tech report uploaded 05:34Z. The deepseekv41-flash-0909 container tag is a day earlier, but that is the image build date, not the weights date. First sweep is same-day — day zero.
• AgentX-only in the overview matrix. Only agentic-traces rows exist; curating it this way keeps the matrix from showing an 8K/1K row that will never fill. Remove the OVERVIEW_MODEL_SCENARIOS entry once a fixed-sequence sweep lands.
• The compare-slug entry is structurally required, not scope creep. INFERENCE_MODEL_SLUGS, MODEL_ROUTES, /rankings and /run all derive from COMPARE_MODEL_SLUGS, and their tests demand coverage for every non-hidden model. As a side effect the model now appears on /compare, /compare-per-dollar, /inference/deepseek-v41-flash, /calculator/…, /historical/… and the sitemap. The hardcoded DESCRIPTION catalog blurbs on /compare and /compare-per-dollar were left alone — say the word if it should be named there.
• OpenRouter id verified, not guessed: deepseek/deepseek-v4.1-flash exists in the live /api/v1/models catalog.

Deliberately not in this PR

• MODEL_ARCHITECTURES — no diagram renders until added; wants verified config.json values.
• The MTP_ENGINE_EXCLUSION that V4-Pro carries. Only vLLM MTP has swept this model, so the guard has nothing to separate yet. It should be added the moment an SGLang MTP run lands on the same SKU.

Test expectation updates

All four are mechanical consequences of one more visible model, not behavior changes: the compare-slug SEO-name map, the overview fixture row order, /rankings 24 → 26 pages, /run 108 → 117 candidates. The two api-route-catalog digests were refreshed after confirming the docs need no copy change — both SUPPORTED_TCO_MODELS and the overview scenario curation are derived from the data, not hand-written into the reference.

Chinese pages

No /zh page authoring needed. Model names, hardware SKUs and framework names stay English per the translation quality bar, and every affected surface (/zh/compare/[slug], /zh/inference/[model], /zh/calculator/[model], …) is a dynamic route that picks the new model up from the same registries.

Verification

bun run typecheck, bun run lint, bun run fmt, and bun run test:unit (5297 passed, 4 skipped) all pass locally.

⚠️ Deploy order matters. Per docs/adding-entities.md, the packages/db and packages/constants changes must be merged and deployed before the next ingest — otherwise dsv41flash rows resolve to null and are silently skipped.

中文说明

将 DeepSeek-V4.1-Flash 注册为仪表板模型，数据来自 SemiAnalysisAI/InferenceX#2961 的 GB300 AgentX sweep —— DB key 为 dsv41flash，HF 路径 deepseek-ai/DeepSeek-V4.1-Flash，在 gb300-nv 上以 FP4 / vLLM / TP4 / MTP 运行，并发 1–128。

需要评审的几个判断

• 单独建桶，而非 dsv4 的小版本更新。 V4.1-Flash 采用 Causal Encoder-Decoder 架构与 CSA2 稀疏注意力，与 V4-Pro 在结构上不同，合并显示会同时误导两者。处理方式与 kimik3（相对 K2 系列）、minimaxm3（相对 M2 系列）一致。
• 标签写 552B，而不是 748B。 HF model card 给出 552B backbone（prefill 激活 8B，decode 激活 16B），另有 196B 的 Engram 条件记忆表。Engram 通过 token 查表稀疏访问，并非每 token 常驻计算，因此不计入总量 —— 与此前排除独立 MTP head 的口径一致。这里特别标注出来，是因为 Qwen3.8-Flash-Next 的先例恰好相反：其 51B n-gram embedding 表是计入 176B 总量的。如果认为 Engram 属于同一类，标签应改为 748B。
• 发布日期 2026-09-10。 依据 Hugging Face 仓库今日 02:17Z 的首次提交，技术报告于 05:34Z 上传。容器 tag 中的 0909 早一天，但那是镜像构建日期而非权重发布日期。首次 sweep 与发布同日，属于 day zero。
• overview 矩阵中标记为仅 AgentX。 目前只有 agentic traces 数据，这样标记可以避免矩阵中出现永远不会填充的 8K/1K 行。待定长序列 sweep 上线后移除该 OVERVIEW_MODEL_SCENARIOS 条目。
• compare-slug 条目是结构性必需项，并非范围扩张。 INFERENCE_MODEL_SLUGS、MODEL_ROUTES、/rankings 与 /run 均由 COMPARE_MODEL_SLUGS 派生，相关测试要求覆盖每个非 hidden 模型。因此该模型会同时出现在 /compare、/compare-per-dollar、/inference/deepseek-v41-flash、/calculator/…、/historical/… 以及 sitemap 中。/compare 与 /compare-per-dollar 中硬编码的 DESCRIPTION 文案未改动，如需在其中列出该模型请告知。
• OpenRouter id 已核实（非猜测）：deepseek/deepseek-v4.1-flash 存在于线上 /api/v1/models 目录中。

本 PR 有意未包含

• MODEL_ARCHITECTURES：缺失时不渲染架构图，需要经核实的 config.json 数值。
• V4-Pro 所带的 MTP_ENGINE_EXCLUSION：目前只有 vLLM MTP 跑过该模型，该保护规则尚无可区分的对象。一旦同一 SKU 上出现 SGLang MTP 运行，应立即补上。

测试预期更新

四处改动都只是"多了一个可见模型"的机械结果，并非行为变更：compare-slug 的 SEO 名称映射、overview fixture 行序、/rankings 24 → 26 个页面、/run 108 → 117 个候选。两处 api-route-catalog digest 是在确认文档无需改动文案后刷新的 —— SUPPORTED_TCO_MODELS 与 overview 场景 curation 都由数据派生，并非手写进 API 参考。

中文页面

无需新增 /zh 页面。按翻译质量标准，模型名称、硬件 SKU 与框架名称保持英文；所有受影响的页面（/zh/compare/[slug]、/zh/inference/[model]、/zh/calculator/[model] 等）都是动态路由，会自动从同一批注册表中读取新模型。

验证

本地 bun run typecheck、bun run lint、bun run fmt、bun run test:unit（5297 通过，4 跳过）全部通过。

⚠️ 注意部署顺序。 依据 docs/adding-entities.md，packages/db 与 packages/constants 的改动必须在下一次 ingest 之前合并并部署，否则 dsv41flash 行会解析为 null 并被静默跳过。

🤖 Generated with Claude Code

---

Note

Medium Risk
Broad but additive registry wiring across app, constants, and DB ETL; wrong deploy order before ingest would drop rows, and AgentX-only curation must stay aligned with compare routing.

Overview
Adds DeepSeek-V4.1-Flash end-to-end as its own model (DB key dsv41flash, compare slug deepseek-v41-flash), separate from DeepSeek-V4-Pro, with labels anchored on the 552B backbone and AgentX-only curation until an 8K/1K sweep exists.

Data & ingest: packages/constants and packages/db ETL map deepseek-ai/DeepSeek-V4.1-Flash → dsv41flash and record release date 2026-09-10. App registries: Model / MODEL_CONFIG, COMPARE_MODEL_SLUGS, compare SSR KNOWN_MODELS, and /overview OVERVIEW_MODEL_SCENARIOS (AgentX row after V4 Pro). Compare behavior: compare-agentx splits editorial featured AgentX models from factual AgentX-only slugs so V4.1 Flash links to /…/agentic without joining the hero ledger or NEW badges; tests lock that agreement with the overview matrix.

Docs & QA: AGENTS.md parameter table and a new mislabel trap for CED/Engram sizing; Cypress overview expectations (8 matrix rows); bumped rankings/run page counts; refreshed api-route-catalog digests after confirming no public API doc copy change.

^{Reviewed by Cursor Bugbot for commit e68f2ae. 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
inferencemax-app	Ready	Preview	Sep 11, 2026 1:15am UTC

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit bfbe3ed. Configure here.}

[Oseltamivir]

Rebased onto master and fixed the E2E failure.

What failed: overview.cy.ts asserts an exact matrix row count. The new model adds one AgentX row, so the count assertion failed with Found '9', expected '8' on chrome/firefox shard 2.

Fix: MATRIX_ROWS updated, and DeepSeek V4.1 Flash 552B added to MODEL_LABELS and to the AgentX-scenario assertion loop — so the new row is actually verified, not just counted.

The rebase also picked up #1114 (DeepSeek V4 Pro single-turn 8k1k retirement), which touches the same constant and the same overview-data.ts review digest. Both resolved together: master's 7 rows + this model's one AgentX row = 8, and the refreshed digest now covers both changes. Neither alters an API parameter or the OverviewPageData shape, so the API reference copy stands.

Verified locally against a fixtures build (E2E_FIXTURES=1): overview.cy.ts 43 passing, plus overview-a11y, sanity, zh-pages, compare-table — 56 passing total. Unit suite 5343 passed / 4 skipped; typecheck, lint, fmt clean.

中文说明

已 rebase 到 master 并修复 E2E 失败。

失败原因： overview.cy.ts 对矩阵行数做精确断言。新模型新增一行 AgentX 记录，因此 chrome/firefox shard 2 报出 Found '9', expected '8'。

修复： 更新 MATRIX_ROWS，并将 DeepSeek V4.1 Flash 552B 加入 MODEL_LABELS 与 AgentX 场景断言循环，使新行被实际校验而非仅计入行数。

rebase 同时合入了 #1114（停用 DeepSeek V4 Pro 单轮 8k1k 场景），该 PR 改动了同一常量与同一份 overview-data.ts review digest。两者一并处理：master 的 7 行加上本 PR 的 1 行 AgentX，合计 8 行；刷新后的 digest 同时覆盖两处改动。两者均未改变 API 参数或 OverviewPageData 结构，API 参考文档无需改动。

本地验证（E2E_FIXTURES=1 fixtures 构建）：overview.cy.ts 43 项通过，另有 overview-a11y、sanity、zh-pages、compare-table，合计 56 项通过。单元测试 5343 通过 / 4 跳过；typecheck、lint、fmt 均通过。

[Oseltamivir]

The remaining red check (E2E chrome shard 3) is a pre-existing failure on master, not caused by this PR.

profit-estimator.cy.ts → "Profit Estimator per GW > opens with the documented defaults and a priced chart" fails at the model-selector step:

AssertionError: Timed out retrying after 6000ms:
Expected to find element: `[role="option"]`, but never found it.

Evidence it is not this PR: I built origin/master in a clean worktree with E2E_FIXTURES=1 and ran the same spec — identical failure, 41 passing / 1 failing. Same result on this branch. It reproduces on both, so the cause arrived with one of the PowerX profit-estimator commits this branch rebased onto (#1115 / #1116 / #1118), not here.

Why this PR cannot affect that selector: offeredModels in ProfitEstimatorDisplay.tsx intersects the tab's model routes with modelsWithAgenticData(...). DeepSeek-V4.1-Flash has no rows in the e2e fixtures, so it is filtered out and the dropdown still offers exactly the same four models. The screenshot confirms the shape of the failure too — the dropdown never opens at all (0 options found), rather than opening with a changed option count.

I have not touched that spec. Someone should open a separate issue against the PowerX changes; I did not want to fold an unrelated fix into this PR.

Everything else is green: both overview shards now pass, along with lint, typecheck, unit tests, component tests, CodeQL and Bugbot.

中文说明

剩余的红色检查项（E2E chrome shard 3）是 master 上已存在的失败，与本 PR 无关。

profit-estimator.cy.ts 中 "Profit Estimator per GW > opens with the documented defaults and a priced chart" 在模型选择器一步失败，报错为找不到 [role="option"] 元素。

证明与本 PR 无关： 我在干净的 worktree 中以 E2E_FIXTURES=1 构建 origin/master 并运行同一个 spec，出现完全相同的失败（41 通过 / 1 失败）；本分支结果一致。两边均可复现，说明问题来自本分支 rebase 时合入的 PowerX 利润估算器改动（#1115 / #1116 / #1118），而非本 PR。

本 PR 不可能影响该选择器： ProfitEstimatorDisplay.tsx 中的 offeredModels 会将该标签页的模型路由与 modelsWithAgenticData(...) 取交集。DeepSeek-V4.1-Flash 在 e2e fixtures 中没有数据，因此会被过滤掉，下拉列表仍然只有原来的四个模型。截图也印证了失败形态：下拉根本没有展开（找到 0 个选项），而不是展开后选项数量发生变化。

我没有改动该 spec。建议针对 PowerX 的改动另开 issue 处理，不宜把无关修复混入本 PR。

其余检查全部通过：两个 overview 分片均已通过，lint、typecheck、单元测试、组件测试、CodeQL 与 Bugbot 也都通过。

[Oseltamivir]

Addressed all three review findings, rebased onto master (now at #1119).

1. AgentX-only models fell through to the 8K/1K default (the Bugbot finding). Fixed by separating the two questions that FEATURED_AGENTX_MODEL_SLUGS was answering at once. That list is editorial — it drives the compare hero ledger and the NEW badge in the /inference selector — but comparisonScenarioForModel was also using it to decide which workload a model has data for. Those coincided until this model.

AGENTX_ONLY_MODEL_SLUGS now feeds the scenario lookup alongside the featured list; the hero and NEW badge still read the featured list alone. So the routing is correct without promoting the model into the hero — that stays a separate product decision.

Added a drift guard: a test pins the compare scenario set against OVERVIEW_MODEL_SCENARIOS, which encodes the same fact for the overview matrix. Curating one without the other now fails a test instead of silently SSR-ing an empty page. The guard lives in the test rather than as a runtime import — ChartControls is a client component, so importing the overview module into compare-agentx.ts would pull the matrix builder into the browser bundle. Both new tests were mutation-checked: reverting the fix fails them.

2. AGENTS.md parameter table. Added the row and a mislabel-trap bullet. The trap is worth having in writing: the active count is two numbers (8B prefill / 16B decode) because of the CED split, the 196B Engram table is excluded from the 552B backbone while Qwen3.8's n-gram table is counted in its total, and HF safetensors metadata reports 485B, matching neither.

3. MODEL_RELEASE_DATES ordering. The map is documented "Newest first" so the weights-predate-sweep invariant stays checkable by eye; 2026-09-10 is the newest date in the table but sat between MiniMax M3 and V4 Pro. Moved to the top.

Considered, not changed: a reviewer read the compare-slug ordering rule ("newer family member leads") as requiring deepseek-v41-flash to precede deepseek-v4. That rule is illustrated with M3 leading M2 and K3 leading K2 — newer flagship over older flagship. V4.1 Flash is a smaller sibling of the current flagship, and leading with it would push V4 Pro down in a list whose first rule is "flagship Chinese-developed models first." Happy to flip it if you read it the other way.

Unit suite 5348 passed / 4 skipped; typecheck, lint, fmt clean.

中文说明

已处理全部三项评审意见，并 rebase 到最新 master（#1119）。

1. 仅有 AgentX 数据的模型回退到 8K/1K 默认值（Bugbot 指出的问题）。 根因是 FEATURED_AGENTX_MODEL_SLUGS 同时承担了两个职责：它本是编辑性列表（驱动对比页 hero 栏与 /inference 选择器的 NEW 标记），却也被 comparisonScenarioForModel 用来判断模型究竟有哪种 workload 数据。在本模型出现之前，两者恰好一致。

现新增 AGENTX_ONLY_MODEL_SLUGS，与 featured 列表共同决定场景；hero 与 NEW 标记仍只读取 featured 列表。因此路由得到修复，而不会把该模型推上 hero —— 后者仍是独立的产品决策。

同时新增漂移守卫：通过测试将对比场景集合与 overview 矩阵中记录同一事实的 OVERVIEW_MODEL_SCENARIOS 对齐，今后只改其一会直接测试失败，而不是静默渲染空页面。守卫放在测试而非运行时导入：ChartControls 是客户端组件，在 compare-agentx.ts 中引入 overview 模块会把矩阵构建逻辑带进浏览器产物。两项新测试均做过变异验证：回退修复后会失败。

2. AGENTS.md 参数表。 已补充该模型行与一条易错说明。这条说明值得写清楚：受 CED 架构拆分影响，激活参数是两个数值（prefill 8B / decode 16B）；196B 的 Engram 表不计入 552B backbone，而 Qwen3.8 的 n-gram 表是计入其总量的；此外 HF safetensors 元数据显示 485B，与两者皆不符。

3. MODEL_RELEASE_DATES 排序。 该表注释要求"最新在前"，以便用肉眼核对"权重日期不得晚于首次 sweep"这一不变式；2026-09-10 是表中最新日期，却位于 MiniMax M3 与 V4 Pro 之间。已移至表首。

评估后未改动： 有评审意见将 compare-slug 的排序规则（"较新的同系列成员在前"）理解为 deepseek-v41-flash 应排在 deepseek-v4 之前。但该规则的示例是 M3 先于 M2、K3 先于 K2，即较新的旗舰先于较旧的旗舰。V4.1 Flash 是当前旗舰的小尺寸同系列成员，若将其前置，会在"中国实验室旗舰模型优先"这一首要规则下把 V4 Pro 往后挤。如果你的理解不同，我可以调整。

单元测试 5348 通过 / 4 跳过；typecheck、lint、fmt 均通过。