DeepSeek V4.1 Flash Provider Benchmark

[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6,"_7":8,"_9":10},"actionData","errors","root",{"_494":495,"_496":497},"routes/_app.x.$handle.$name",{"_111":110,"_38":52,"_119":97,"_33":120,"_13":14,"_121":-5,"_11":122,"_123":124,"_125":126,"_127":128,"_129":130,"_26":-5,"_131":132,"_133":134,"_135":31,"_136":137,"_138":66,"_139":-5,"_140":-5,"_141":142},"routes/_app.x.$handle.$name.branch.$branch.version.$version.diff",{"_11":12,"_13":14,"_15":16,"_17":18,"_19":-5},"project",{"_50":51,"_38":52,"_45":53,"_54":-5,"_55":31,"_56":57,"_58":57,"_59":60,"_61":-5,"_19":62,"_63":64,"_65":66,"_67":-5,"_68":-5,"_69":70,"_71":72,"_73":74,"_75":76,"_77":78,"_79":-5,"_80":81,"_82":81,"_83":84,"_85":84,"_86":84,"_87":81,"_88":81,"_89":90,"_91":92,"_93":94,"_95":31},"version",41,"changes",[25],"aiChatInfo",{"_20":21,"_22":-5,"_23":-5,"_24":-5},"userId","aiChatMessageId","ad3e6c4d-727e-484b-9fc6-62dd5af373b2","aiChatThreadId","aiChatThreadUserId","aiChatThreadTitle",{"_26":27,"_28":29,"_30":31,"_32":31,"_33":34,"_35":34},"file",{"_36":37,"_38":39,"_40":41,"_42":44,"_45":46,"_47":14,"_48":49,"_30":31},"previous",{"_36":37,"_38":39,"_40":41,"_42":43},"deleted",false,"deletedHere","path","benchmark/prompt.ts","latestPath","fileId","01a08da5-899f-73ab-baf1-8c16de22b830","name","prompt.ts","fileType","script","content","import type { ModelMessage } from \"npm:ai@6.0.168\";\n\n// Bump this whenever the messages or metric definitions change.\n// Runs with a different version stay available as raw measurements but are not\n// pooled into the current provider comparison.\nexport const BENCHMARK_VERSION = \"static-multiturn-v2\";\n\nexport const BENCHMARK_MESSAGES = [\n {\n role: \"system\",\n content: `\nYou are reviewing a TypeScript service that coordinates background jobs. Treat every\nmessage before the final request as durable project context. Preserve public API\nbehavior, prefer explicit types, use parameterized database queries, avoid hidden\nglobal state, and explain operational tradeoffs. The service runs on a serverless\nruntime with no writable filesystem. It can call HTTP services and a scoped SQLite\ndatabase. Jobs may be retried, so every write must be idempotent. Timestamps are ISO\n8601 UTC strings. Errors must be recorded without discarding successful work from\nother providers. Readers should be able to distinguish an attempted operation from\na completed one. Never claim that an external request succeeded unless its response\nwas observed. Keep recommendations concise and actionable.\n\nThe architecture has three layers. The HTTP layer is read only and returns normalized\nrecords. The worker layer performs expensive outbound calls and checkpoints each\nattempt. The persistence layer owns schema creation and all SQL. A run contains many\nprovider attempts. Each attempt records latency, time to first text, token counts,\ncache reads, cache writes, cost, a short response preview, provider metadata, and an\nerror when present. The UI aggregates only successful attempts for performance\nfigures while displaying failures separately. A provider can succeed on some attempts\nand fail on others. Missing metrics must remain null instead of silently becoming\nzero. The code should be easy for another engineer to inspect, remix, and operate.\n`.trim(),\n },\n {\n role: \"user\",\n content: `We need an idempotent schema for benchmark runs and their individual\nprovider attempts. Runs have a stable text id, model id, status, start/end timestamps,\nconfiguration, and endpoint snapshot. Attempts are unique by run, provider, and\nattempt number. We need to resume after interruption without duplicating work.`,\n },\n {\n role: \"assistant\",\n content: `Use a benchmark_runs table for lifecycle state and a\nbenchmark_measurements table with a composite primary key of\n(run_id, provider, attempt). Store nullable observed metrics as nullable numeric\ncolumns and keep raw provider data as JSON text. Create an index on run start time and\non measurement run_id. Insert each completed measurement once and treat a conflict as\nan already-checkpointed attempt. Resume by reading the composite keys already present\nand scheduling only missing attempts.`,\n },\n {\n role: \"user\",\n content: `The public dashboard needs an aggregate provider comparison and a\nfilterable table of individual measurements. It must never expose the gateway\ncredential or provide a mutation that starts work. How should the API be shaped?`,\n },\n {\n role: \"assistant\",\n content: `Expose read-only GET endpoints for aggregate results and paginated\nmeasurements. Use parameterized SQL for filters, clamp pagination server side, and do\nnot define POST, PUT, PATCH, or DELETE routes. The benchmark worker can read the\ncredential from its environment because that value never enters an HTTP response or\nbrowser bundle. Return no-store headers while results may still change.`,\n },\n {\n role: \"user\",\n content: `Provider calls stream text. We want time to first text, total duration,\noutput throughput, usage details, cache tokens, and gateway cost. Some streams fail\nbefore producing text and others fail after partial output.`,\n },\n {\n role: \"assistant\",\n content: `Record a monotonic start time immediately before stream creation.\nSet TTFT only when the first text delta arrives. Consume the complete stream before\ndeciding whether usable text was returned, while storing only a bounded preview.\nCompute throughput only when output tokens and a positive interval after TTFT are\navailable. Wrap each attempt independently so a failure becomes an error measurement\nwith its observed timing and does not stop other providers. Keep missing usage and\nprice values null.`,\n },\n {\n role: \"user\",\n content: `There are many providers and multiple attempts per provider. Cache\nbehavior depends on repeated prompts, but a completely serial run is slow. What\nconcurrency model preserves the useful signal?`,\n },\n {\n role: \"assistant\",\n content: `Send the exact same messages for every attempt. Run attempts\nsequentially within each provider so attempt two follows attempt one. Process a small\nnumber of different providers concurrently because their caches are independent. Use\na bounded provider worker pool and checkpoint every attempt immediately. This shortens\nwall clock time while preserving the within-provider warmup sequence.`,\n },\n {\n role: \"user\",\n content: `The endpoint catalogue includes advertised caching, pricing, recent\nlatency, and throughput. These fields can be absent or change over time. How should\nwe preserve and display them?`,\n },\n {\n role: \"assistant\",\n content: `Capture the full endpoint records at the start of each benchmark and\nstore that snapshot with the run. Copy the provider record onto each measurement so\nraw exports remain understandable. Render absent values as unknown, label catalogue\nvalues as Vercel metadata, and keep observed metrics separate from advertised metrics.\nAn interrupted run must resume against its original endpoint snapshot.`,\n },\n {\n role: \"user\",\n content: `The result page needs to compare providers without suggesting false\nprecision. Four attempts per run is a small sample and failures are part of the\noutcome.`,\n },\n {\n role: \"assistant\",\n content: `Pool comparable completed runs. Show reliability, median and tail TTFT,\nmedian and slow-tail post-TTFT throughput, eligible cache hit rate, and observed cost\nper million input plus output tokens. Keep every column sortable and retain individual\nattempts on a separate measurements page. Never hide providers with zero successes.`,\n },\n {\n role: \"user\",\n content: `Another engineer should be able to change the model or scope and run\nthis from the code view. They should not need to understand the UI implementation.`,\n },\n {\n role: \"assistant\",\n content: `Keep a small benchmark/config.ts containing the benchmark version,\nmodel id, run count, output cap, timeout, concurrency, and an optional provider\nallowlist. Put the fixed conversation in a separate prompt module. Make the scheduled\ninterval the obvious code-level entrypoint: edit the config, open scheduled.ts, and\nclick Run. Public HTTP routes must remain read only.`,\n },\n {\n role: \"user\",\n content: `We need a handoff checklist for operating the benchmark after the\noriginal author leaves. Include recovery from partial runs and provider failures.`,\n },\n {\n role: \"assistant\",\n content: `Document the credential name without its value, the scheduled/manual\ncode entrypoint, the database tables, and the read-only routes. Explain that the same\nbenchmark configuration resumes the latest running record and skips persisted\nattempts. A completed run includes every scheduled attempt, whether successful or\nfailed. Preserve provider error rows and start a later run rather than rewriting a\ncompleted run. Verify the page, API, and logs after edits.`,\n },\n {\n role: \"user\",\n content: `In exactly two sentences, name the most important invariant in this\narchitecture and one metric that would reveal a broken implementation. Do not use a\nheading or bullet list.`,\n },\n] satisfies ModelMessage[];\n","import type { ModelMessage } from \"npm:ai@6.0.168\";\n\n// Bump this whenever the messages or metric definitions change.\n// Runs with a different version stay available as raw measurements but are not\n// pooled into the current provider comparison.\n//\n// v3: the final question is pure recall so a reasoning model spends ~80–160\n// reasoning tokens (measured on novita/particle/deepseek) and still fits\n// under the 300-token output cap. The previous open-ended question burned\n// 100–370 reasoning tokens and often produced no visible text at all.\nexport const BENCHMARK_VERSION = \"static-multiturn-v3\";\n\nexport const BENCHMARK_MESSAGES = [\n {\n role: \"system\",\n content: `\nYou are reviewing a TypeScript service that coordinates background jobs. Treat every\nmessage before the final request as durable project context. Preserve public API\nbehavior, prefer explicit types, use parameterized database queries, avoid hidden\nglobal state, and explain operational tradeoffs. The service runs on a serverless\nruntime with no writable filesystem. It can call HTTP services and a scoped SQLite\ndatabase. Jobs may be retried, so every write must be idempotent. Timestamps are ISO\n8601 UTC strings. Errors must be recorded without discarding successful work from\nother providers. Readers should be able to distinguish an attempted operation from\na completed one. Never claim that an external request succeeded unless its response\nwas observed. Keep recommendations concise and actionable.\n\nThe architecture has three layers. The HTTP layer is read only and returns normalized\nrecords. The worker layer performs expensive outbound calls and checkpoints each\nattempt. The persistence layer owns schema creation and all SQL. A run contains many\nprovider attempts. Each attempt records latency, time to first text, token counts,\ncache reads, cache writes, cost, a short response preview, provider metadata, and an\nerror when present. The UI aggregates only successful attempts for performance\nfigures while displaying failures separately. A provider can succeed on some attempts\nand fail on others. Missing metrics must remain null instead of silently becoming\nzero. The code should be easy for another engineer to inspect, remix, and operate.\n`.trim(),\n },\n {\n role: \"user\",\n content: `We need an idempotent schema for benchmark runs and their individual\nprovider attempts. Runs have a stable text id, model id, status, start/end timestamps,\nconfiguration, and endpoint snapshot. Attempts are unique by run, provider, and\nattempt number. We need to resume after interruption without duplicating work.`,\n },\n {\n role: \"assistant\",\n content: `Use a benchmark_runs table for lifecycle state and a\nbenchmark_measurements table with a composite primary key of\n(run_id, provider, attempt). Store nullable observed metrics as nullable numeric\ncolumns and keep raw provider data as JSON text. Create an index on run start time and\non measurement run_id. Insert each completed measurement once and treat a conflict as\nan already-checkpointed attempt. Resume by reading the composite keys already present\nand scheduling only missing attempts.`,\n },\n {\n role: \"user\",\n content: `The public dashboard needs an aggregate provider comparison and a\nfilterable table of individual measurements. It must never expose the gateway\ncredential or provide a mutation that starts work. How should the API be shaped?`,\n },\n {\n role: \"assistant\",\n content: `Expose read-only GET endpoints for aggregate results and paginated\nmeasurements. Use parameterized SQL for filters, clamp pagination server side, and do\nnot define POST, PUT, PATCH, or DELETE routes. The benchmark worker can read the\ncredential from its environment because that value never enters an HTTP response or\nbrowser bundle. Return no-store headers while results may still change.`,\n },\n {\n role: \"user\",\n content: `Provider calls stream text. We want time to first text, total duration,\noutput throughput, usage details, cache tokens, and gateway cost. Some streams fail\nbefore producing text and others fail after partial output.`,\n },\n {\n role: \"assistant\",\n content: `Record a monotonic start time immediately before stream creation.\nSet TTFT only when the first text delta arrives. Consume the complete stream before\ndeciding whether usable text was returned, while storing only a bounded preview.\nCompute throughput only when output tokens and a positive interval after TTFT are\navailable. Wrap each attempt independently so a failure becomes an error measurement\nwith its observed timing and does not stop other providers. Keep missing usage and\nprice values null.`,\n },\n {\n role: \"user\",\n content: `There are many providers and multiple attempts per provider. Cache\nbehavior depends on repeated prompts, but a completely serial run is slow. What\nconcurrency model preserves the useful signal?`,\n },\n {\n role: \"assistant\",\n content: `Send the exact same messages for every attempt. Run attempts\nsequentially within each provider so attempt two follows attempt one. Process a small\nnumber of different providers concurrently because their caches are independent. Use\na bounded provider worker pool and checkpoint every attempt immediately. This shortens\nwall clock time while preserving the within-provider warmup sequence.`,\n },\n {\n role: \"user\",\n content: `The endpoint catalogue includes advertised caching, pricing, recent\nlatency, and throughput. These fields can be absent or change over time. How should\nwe preserve and display them?`,\n },\n {\n role: \"assistant\",\n content: `Capture the full endpoint records at the start of each benchmark and\nstore that snapshot with the run. Copy the provider record onto each measurement so\nraw exports remain understandable. Render absent values as unknown, label catalogue\nvalues as Vercel metadata, and keep observed metrics separate from advertised metrics.\nAn interrupted run must resume against its original endpoint snapshot.`,\n },\n {\n role: \"user\",\n content: `The result page needs to compare providers without suggesting false\nprecision. Four attempts per run is a small sample and failures are part of the\noutcome.`,\n },\n {\n role: \"assistant\",\n content: `Pool comparable completed runs. Show reliability, median and tail TTFT,\nmedian and slow-tail post-TTFT throughput, eligible cache hit rate, and observed cost\nper million input plus output tokens. Keep every column sortable and retain individual\nattempts on a separate measurements page. Never hide providers with zero successes.`,\n },\n {\n role: \"user\",\n content: `Another engineer should be able to change the model or scope and run\nthis from the code view. They should not need to understand the UI implementation.`,\n },\n {\n role: \"assistant\",\n content: `Keep a small benchmark/config.ts containing the benchmark version,\nmodel id, run count, output cap, timeout, concurrency, and an optional provider\nallowlist. Put the fixed conversation in a separate prompt module. Make the scheduled\ninterval the obvious code-level entrypoint: edit the config, open scheduled.ts, and\nclick Run. Public HTTP routes must remain read only.`,\n },\n {\n role: \"user\",\n content: `We need a handoff checklist for operating the benchmark after the\noriginal author leaves. Include recovery from partial runs and provider failures.`,\n },\n {\n role: \"assistant\",\n content: `Document the credential name without its value, the scheduled/manual\ncode entrypoint, the database tables, and the read-only routes. Explain that the same\nbenchmark configuration resumes the latest running record and skips persisted\nattempts. A completed run includes every scheduled attempt, whether successful or\nfailed. Preserve provider error rows and start a later run rather than rewriting a\ncompleted run. Verify the page, API, and logs after edits.`,\n },\n {\n role: \"user\",\n content: `In exactly two sentences, name the most important invariant in this\narchitecture and one metric that would reveal a broken implementation. Do not use a\nheading or bullet list.`,\n },\n] satisfies ModelMessage[];\n","createdAt",["D",1789486687088],"branchVersion","branchId","0cb5a767-9630-4119-a265-8c9f5f633c9a","id","1ea4658b-1977-4a2c-8356-9d6ab6609397","deepseek-v41-flash-benchmark",["D",1789082831202],"deletedAt","legacyEnvVars","privacy","public","httpPrivacy","description","Run and compare AI Gateway providers for deepseek-v4.1-flash","imageUrl","a0bf3b31-15a5-4d5c-880e-4b1e22c9bc18","pinnedFileId","01a08da5-899f-73ab-baf1-f0227ea1ff9b","pinPreview",true,"pinMigrated","screenshotVersion","likeCount",0,"orgId","e81e7335-6d4c-4764-a902-d0962bbfa9f7","tursoName","1ea4658b19774a2c83569d6ab6609397","tursoHostname","1ea4658b19774a2c83569d6ab6609397-val-town.aws-us-east-2.turso.io","tursoMeasuredAt",["D",1789518676100],"pinned","userPerm","r","valPerm","blobPerm","rw","emailPerm","sqlitePerm","projectPerm","telemetryPerm","user",{"_111":110,"_118":114},"org",{"_111":110,"_112":-5,"_113":114,"_19":62,"_115":116,"_117":31},"branch",{"_13":96,"_38":97,"_50":49,"_98":51,"_99":100,"_101":102,"_103":57,"_104":105,"_106":107,"_108":52,"_109":110},"isOwned",81,"main","projectId","forkedBranchId","24cb14b4-1851-4548-adc6-43a574860a3d","forkedBranchVersion",322,"forkedProjectPrivacy","forkedProjectId","d055324c-28f9-4529-80dd-e9bbfbdbeb65","forkedProjectOrgId","ba7bb5fb-1abb-4ffb-8f3d-7c244782e5f7","projectName","userHandle","stevekrouse","handle","displayName","avatar","https://images.clerk.dev/uploaded/img_2PqHa2Gsy93xQrjh2w78Xu0cChW.jpeg","tier","pro","canRestrictVals","profileImageUrl","branchName","","adoptOrgId",{"_50":51,"_38":52,"_45":489,"_54":-5,"_55":31,"_56":57,"_58":57,"_59":60,"_61":-5,"_19":62,"_63":64,"_65":66,"_67":-5,"_68":-5,"_69":70,"_71":72,"_73":74,"_75":76,"_77":490,"_79":-5,"_80":81,"_82":81,"_83":84,"_85":84,"_86":84,"_87":81,"_88":81,"_89":491,"_91":492,"_93":493,"_95":31},"tree",[218,229,234,249,251,252,253],"files",[218,219,220,221,222,223,224,225,226,227,228,229,230,231,232,233,234,235,236,237,238,239,240,241,242,243,244,245,246,247,248,249,250,251,252,253],"counts",{"_208":70,"_209":164,"_210":70,"_211":212,"_213":214},"contents",[154,155,156,157,158,159,160],"upstreamProject",{"_50":105,"_38":150,"_59":151,"_61":-5,"_56":57,"_89":152},"upstreamBranch",{"_50":100,"_45":143,"_144":145,"_38":97,"_98":105,"_13":146,"_99":-5,"_101":-5,"_147":148,"_54":-5,"_95":31,"_56":57,"_109":149,"_71":107,"_108":150},"emptyEnvVars","flags",[],"hasHttpFiles","grants","orgs","customDomains",{},["D",1787848294548],"updatedAt",["D",1789083925414],338,"lockfileKey","1789083928454-esm.town_v_valdottown_glm53-flash-benchmark@338-main","valdottown","glm53-flash-benchmark","Run and compare AI Gateway providers",{"_111":149,"_118":153},"https://img.clerk.com/eyJ0eXBlIjoicHJveHkiLCJzcmMiOiJodHRwczovL2ltYWdlcy5jbGVyay5kZXYvdXBsb2FkZWQvaW1nXzJyWnQ5cFE0c3pleVRCSEVTWmxzaXJsTUlGciJ9",{"_50":203,"_38":204,"_36":205,"_163":-5,"_48":49,"_47":164,"_165":66,"_45":206,"_13":70,"_40":191,"_42":-5,"_30":31,"_169":-5,"_33":207},{"_50":198,"_38":199,"_36":200,"_163":-5,"_48":49,"_47":164,"_165":66,"_45":201,"_13":70,"_40":191,"_42":-5,"_30":31,"_169":-5,"_33":202},{"_50":193,"_38":194,"_36":195,"_163":-5,"_48":49,"_47":164,"_165":66,"_45":196,"_13":70,"_40":191,"_42":-5,"_30":31,"_169":-5,"_33":197},{"_50":187,"_38":188,"_36":189,"_163":-5,"_48":49,"_47":164,"_165":66,"_45":190,"_13":70,"_40":191,"_42":-5,"_30":31,"_169":-5,"_33":192},{"_50":180,"_38":181,"_36":182,"_163":-5,"_48":49,"_47":183,"_165":31,"_45":184,"_13":177,"_40":26,"_42":185,"_30":31,"_169":-5,"_33":186},{"_50":172,"_38":173,"_36":174,"_163":-5,"_48":49,"_47":175,"_165":31,"_45":176,"_13":177,"_40":26,"_42":178,"_30":31,"_169":-5,"_33":179},{"_50":161,"_38":162,"_36":64,"_163":-5,"_48":49,"_47":164,"_165":66,"_45":166,"_13":70,"_40":167,"_42":168,"_30":31,"_169":170,"_33":171},"8ad4c023-63e7-4774-ae26-ef9aa9990c1f","index.ts","parentId",1,"latest",["D",1789082831202],"http","import { parseVal, serveImmutableFile } from \"https://esm.town/v/std/utils/index.ts\";\nimport { Hono } from \"npm:hono@4\";\nimport { FULL_BENCHMARK_CONFIG } from \"./benchmark/config.ts\";\nimport { listMeasurements } from \"./db/measurements.ts\";\nimport { getBenchmarkResults } from \"./db/results.ts\";\nimport { Root } from \"./frontend/root.tsx\";\nimport type {\n MeasurementPublicStatus,\n MeasurementSortDirection,\n MeasurementSortKey,\n} from \"./shared/types.ts\";\n\nconst MEASUREMENT_SORT_KEYS = new Set([\n \"runStartedAt\",\n \"provider\",\n \"attempt\",\n \"status\",\n \"startedAt\",\n \"ttftMs\",\n \"durationMs\",\n \"tokensPerSecond\",\n \"inputTokens\",\n \"outputTokens\",\n \"cacheReadTokens\",\n \"cacheWriteTokens\",\n \"costUsd\",\n \"message\",\n]);\nconst app = new Hono();\n\nfunction measurementSortKey(value: string | undefined): MeasurementSortKey {\n const key = value as MeasurementSortKey | undefined;\n return key && MEASUREMENT_SORT_KEYS.has(key) ? key : \"startedAt\";\n}\n\nfunction measurementStatus(\n value: string | undefined,\n): MeasurementPublicStatus | undefined {\n return value === \"success\" || value === \"error\" ? value : undefined;\n}\n\nfunction positiveInteger(\n value: string | undefined,\n fallback: number,\n maximum: number,\n): number {\n const parsed = Number(value);\n return Number.isInteger(parsed) && parsed > 0\n ? Math.min(parsed, maximum)\n : fallback;\n}\n\nfunction nonNegativeInteger(\n value: string | undefined,\n fallback: number,\n maximum: number,\n): number {\n const parsed = Number(value);\n return Number.isSafeInteger(parsed) && parsed >= 0\n ? Math.min(parsed, maximum)\n : fallback;\n}\n\napp.get(\"/\", (c) => c.html(Root()));\napp.get(\"/measurements\", (c) => c.html(Root()));\n\napp.get(\"/__immutable/*\", (c) => serveImmutableFile(c.req.path));\n\napp.get(\"/api/results\", async (c) => {\n const results = await getBenchmarkResults(FULL_BENCHMARK_CONFIG.modelId);\n c.header(\"Cache-Control\", \"no-store\");\n c.header(\"X-Content-Type-Options\", \"nosniff\");\n return c.json({ results });\n});\n\napp.get(\"/api/measurements\", async (c) => {\n const attemptValue = c.req.query(\"attempt\");\n const attempt = positiveInteger(attemptValue, 0, 100);\n const direction: MeasurementSortDirection =\n c.req.query(\"direction\") === \"asc\" ? \"asc\" : \"desc\";\n\n const measurements = await listMeasurements({\n modelId: FULL_BENCHMARK_CONFIG.modelId,\n provider: c.req.query(\"provider\") || undefined,\n status: measurementStatus(c.req.query(\"status\")),\n attempt: attempt > 0 ? attempt : undefined,\n runId: c.req.query(\"run\") || undefined,\n sortKey: measurementSortKey(c.req.query(\"sort\")),\n direction,\n limit: positiveInteger(c.req.query(\"limit\"), 50, 100),\n offset: nonNegativeInteger(c.req.query(\"offset\"), 0, 10_000_000),\n });\n\n c.header(\"Cache-Control\", \"no-store\");\n c.header(\"X-Content-Type-Options\", \"nosniff\");\n return c.json({ measurements });\n});\n\napp.get(\"/source\", (c) => c.redirect(parseVal().links.self.val));\n\napp.onError((err, c) => {\n console.error(\n \"Benchmark request failed:\",\n err instanceof Error ? err.message : String(err),\n );\n c.header(\"Cache-Control\", \"no-store\");\n c.header(\"X-Content-Type-Options\", \"nosniff\");\n return c.req.path.startsWith(\"/api/\")\n ? c.json({ error: \"The request could not be completed.\" }, 500)\n : c.text(\"The page could not be loaded.\", 500);\n});\n\nexport default app.fetch;\n","label","deepseek-v41-flash",[162],"1e860068-9acd-475e-9ae5-ea2506a5bcec","README.md","01a08da5-899f-73ab-baf1-ef75ca3ea3d7",35,["D",1789484890481],3,"# DeepSeek V4.1 Flash Gateway Benchmark\n\nA public, read-only Val Town page that compares providers serving\n`deepseek/deepseek-v4.1-flash` through Vercel AI Gateway.\n\n## How results work\n\nEvery provider attempt is one row in the val-scoped\n`benchmark_measurements` SQLite table. The default provider comparison pools\nmeasurements from every completed run for the model. Each run keeps its\nbenchmark version in SQLite so methodology changes remain auditable without\nremoving older completed measurements from the aggregate.\n\nBefore each run, the worker fetches Vercel's live endpoint catalog. It\nbenchmarks every returned provider and stores a normalized provider snapshot.\nProviders newly added by Vercel are included automatically. An interrupted run\nresumes against the exact provider set it started with.\n\nEach provider receives the same fixed multi-turn message array four times.\nAttempts for one provider run sequentially to preserve the cache warmup signal,\nwhile different providers run concurrently. Gateway automatic caching is\nenabled and each request is restricted to the provider being measured.\n\n## Routing strategies\n\nAlongside the solo providers (`only: [provider]`), the run also benchmarks\nfive synthetic \"strategy\" endpoints defined in `benchmark/strategies.ts`. They\nexercise the gateway's own routing instead of pinning a provider:\n\n| Row | `providerOptions.gateway` |\n|-----|---------------------------|\n| Gateway default | no hints — gateway picks by recent uptime + latency |\n| Sort by TTFT | `sort: \"ttft\"` |\n| Sort by throughput | `sort: \"tps\"` |\n| Handpicked only | `only: [modal, wafer, particle]`, `order: [modal, wafer, particle]` |\n| Prefer handpicked | `order: [modal, wafer, particle]` — falls back to any other provider |\n\nEvery strategy also sends `zeroDataRetention: true` and\n`disallowPromptTraining: true`, so the gateway only routes to providers with\nno-store / no-train agreements. Those flags are shown per provider in the\n**No store** / **No train** columns, read from `has_zdr` / `has_no_training`\nin Vercel's endpoint catalogue.\n\nFor strategy rows the table also shows which provider the gateway actually\nserved from (`providerMetadata.gateway.routing.finalProvider`) and how many\nsuccessful requests needed a failover (`totalProviderAttemptCount > 1`). TTFT\nfor a strategy row therefore includes any retry time — the true cost of\nautorouting.\n\n## Known caveat: reasoning burns the token budget\n\n`deepseek-v4.1-flash` is a reasoning model and the benchmark caps output at\n300 tokens. Every `no_text` error has exactly 300 output tokens: the provider\nenabled reasoning by default, the model spent the whole budget thinking, and\nnever emitted text. Providers that don't emit reasoning (wafer, modal,\ndeepinfra) look ~99% reliable; ones that do (novita, deepseek, gmicloud,\nparticle) look 75–87%. It also inflates tokens/sec for reasoning providers,\nbecause all 300 tokens are divided by the tiny window after first text.\nFixing this (higher cap, or disabling reasoning) is a benchmark-version change.\n\nThe sortable provider table reports:\n\n- Reliability across all attempts.\n- TTFT p50 and p95 across successful responses.\n- p50 and p5 observed output tokens per second after first text.\n- Cache hits among eligible repeated successful responses.\n- Observed cost per million input plus output tokens.\n\nTokens/sec is output tokens divided by `duration - TTFT`. It measures how\nquickly text reaches this client after the first text delta. Providers that\nbuffer output into a few large chunks can therefore appear unusually fast.\np50 is the median, TTFT p95 is the slow latency tail, and tokens/sec p5 is the\nslow throughput tail.\n\n$/MTok is total observed cost divided by total input plus output tokens,\nmultiplied by one million. Cached tokens remain part of input tokens, while\ntheir discount is reflected in observed cost.\n\nThe measurements page exposes Success and Error. Success is based on the\ncomplete consumed text stream, independent of the bounded private response\npreview. Error rows show a short, redacted message; timeouts include the\nconfigured 75 second limit. Raw response text and provider metadata remain\nprivate.\n\n## Schedule and recovery\n\n`benchmark/scheduled.ts` runs every 15 minutes in UTC. Requests have a\n75 second timeout and automatic SDK retries are disabled.\n\nA renewable SQLite lease prevents scheduled and manual starts from\noverlapping. The lease can be reclaimed after five minutes without a renewal,\nso an interrupted invocation does not block several later intervals.\nIndividual attempts are checkpointed, and a later invocation resumes the same\nbenchmark version without duplicating stored measurements.\n\nTo run manually, edit `benchmark/config.ts` if needed, open\n`benchmark/scheduled.ts` in Val Town, and click **Run**. Public visitors cannot\nstart model requests from the HTTP app.\n\n## Access\n\nThe Gateway credential is stored as the encrypted\n`VERCEL_AI_GATEWAY_API_KEY` environment variable. The HTTP app exposes only\nread-only routes.\n\n## Structure\n\n- `benchmark/prompt.ts` — fixed messages and benchmark version.\n- `benchmark/attempt.ts` — one streamed AI SDK request.\n- `benchmark/measurement.ts` — pure observation-to-row conversion.\n- `benchmark/endpoints.ts` — live Gateway provider discovery.\n- `benchmark/strategies.ts` — synthetic routing-strategy endpoints.\n- `benchmark/lease.ts` — renewable execution lease.\n- `benchmark/execute.ts` — benchmark orchestration.\n- `db/` — val-scoped SQLite schema, writes, provider snapshots, and aggregation.\n- `frontend/` — React interface styled with Twind.\n- `shared/` — shared benchmark and result types.\n",[173],"debdb676-bc1a-41da-9aed-3b08c5132f69","AGENTS.md","81dba77c-ad6f-11f1-bc49-1607ee4eb77e",38,["D",1789484890661],"# deepseek-v41-flash-benchmark\n\nGateway provider benchmark for `deepseek/deepseek-v4.1-flash` through Vercel AI\nGateway. Remixed from `valdottown/glm53-flash-benchmark` with the model id and\nUI copy switched to DeepSeek. Diverges from upstream by also benchmarking gateway\nrouting strategies (`benchmark/strategies.ts`) and showing ZDR / no-train flags.\n\nGotchas:\n- `no_text` errors = reasoning ate the 300-token `maxOutputTokens` budget (see README caveat).\n- Routing metadata lives in `provider_metadata_json` → `$.gateway.routing.{finalProvider,totalProviderAttemptCount,planningReasoning,sort}`. Only present on successful attempts.\n- Request-level `zeroDataRetention` requires a Vercel Pro/Enterprise plan — confirmed working on this key.\n\n## File map\n\n| Path | Type | Purpose |\n|------|------|---------|\n| `index.ts` | http | Hono read-only app (`/`, `/measurements`, `/api/results`, `/api/measurements`) |\n| `benchmark/config.ts` | script | **Model + run config** — change `modelId` here |\n| `benchmark/prompt.ts` | script | Fixed multi-turn messages, `BENCHMARK_VERSION` |\n| `benchmark/attempt.ts` | script | One streamed AI SDK request |\n| `benchmark/measurement.ts` | script | Pure observation-to-row conversion |\n| `benchmark/endpoints.ts` | script | Live Gateway provider discovery |\n| `benchmark/strategies.ts` | script | Synthetic `gateway:*` routing-strategy endpoints (ZDR + no-train flags, handpicked set) |\n| `benchmark/lease.ts` | script | Renewable execution lease |\n| `benchmark/execute.ts` | script | Benchmark orchestration |\n| `benchmark/scheduled.ts` | interval | Cron entrypoint — runs every 15 min UTC |\n| `db/schema.ts` | script | SQLite schema + `sqlite` export |\n| `db/queries.ts` | script | Run/measurement writes and lookups |\n| `db/results.ts` | script | Aggregate provider comparison query |\n| `db/measurements.ts` | script | Paginated measurements listing |\n| `frontend/root.tsx` | script | HTML shell + title/meta |\n| `frontend/index.tsx` | script | Client entrypoint, page switch |\n| `frontend/components/App.tsx` | script | Model header + provider table view |\n| `frontend/components/ProviderTable.tsx` | script | Sortable provider table (used twice: strategies + providers) |\n| `frontend/components/StrategyCells.tsx` | script | No store / No train ✓ cells, routing summary under strategy names |\n| `frontend/components/MeasurementsPage.tsx` | script | Measurements page |\n| `frontend/components/Sparkline.tsx` | script | History sparklines |\n| `frontend/components/InfoPopover.tsx` | script | Column explainers |\n| `frontend/{providers,history,format}.ts` | script | Provider names, hourly series, formatters |\n| `frontend/favicon.svg` | file | Favicon (still the GLM-era glyph) |\n| `shared/types.ts` | script | Shared benchmark/result types |\n| `README.md` | file | Overview + schedule docs |\n\n## Credential\n\n`VERCEL_AI_GATEWAY_API_KEY` (scoped env var). HTTP app is read-only.\n",[181],"46eb97b6-37cb-403b-8d01-53b8d6528ad3","shared","01a08da5-899f-73ab-baf1-e8049921b5fd",["D",1789082831202],"directory",[188],"88251fa9-177a-4d3b-b712-fbb24a0ddaf6","frontend","01a08da5-899f-73ab-baf1-e025e2c5cf8b",["D",1789082831202],[194],"e6dde3be-d02c-487f-9b40-5146d7076871","db","01a08da5-899f-73ab-baf1-a881a20f339e",["D",1789082831202],[199],"7db55bd3-dcde-471f-a4e0-24c4c57a7efe","benchmark","01a08da5-899f-73ab-baf1-9427bec3af29",["D",1789082831202],[204],"pulls","branches","forks","envVars",{"_215":164,"_217":70},"connections",{"_215":70,"_216":70},"count","unfilledCount","emptyCount",{"_50":203,"_38":204,"_36":205,"_163":-5,"_48":49,"_47":164,"_165":66,"_45":486,"_13":70,"_40":191,"_42":-5,"_30":31,"_169":-5,"_255":487,"_33":488},{"_50":478,"_38":479,"_36":480,"_163":205,"_48":49,"_47":481,"_165":31,"_45":482,"_13":269,"_40":41,"_42":483,"_30":31,"_169":-5,"_255":484,"_33":485},{"_50":471,"_38":472,"_36":473,"_163":205,"_48":49,"_47":269,"_165":31,"_45":474,"_13":164,"_40":41,"_42":475,"_30":31,"_169":-5,"_255":476,"_33":477},{"_50":464,"_38":465,"_36":466,"_163":205,"_48":49,"_47":164,"_165":66,"_45":467,"_13":70,"_40":41,"_42":468,"_30":31,"_169":-5,"_255":469,"_33":470},{"_50":456,"_38":457,"_36":458,"_163":205,"_48":49,"_47":459,"_165":66,"_45":460,"_13":269,"_40":41,"_42":461,"_30":31,"_169":-5,"_255":462,"_33":463},{"_50":449,"_38":450,"_36":451,"_163":205,"_48":49,"_47":164,"_165":66,"_45":452,"_13":70,"_40":41,"_42":453,"_30":31,"_169":-5,"_255":454,"_33":455},{"_50":442,"_38":443,"_36":444,"_163":205,"_48":49,"_47":164,"_165":31,"_45":445,"_13":70,"_40":41,"_42":446,"_30":31,"_169":-5,"_255":447,"_33":448},{"_50":434,"_38":435,"_36":436,"_163":205,"_48":49,"_47":437,"_165":31,"_45":438,"_13":164,"_40":41,"_42":439,"_30":31,"_169":-5,"_255":440,"_33":441},{"_50":430,"_38":39,"_36":37,"_163":205,"_48":49,"_47":14,"_165":31,"_45":431,"_13":164,"_40":41,"_42":44,"_30":31,"_169":-5,"_255":432,"_33":433},{"_50":422,"_38":423,"_36":424,"_163":205,"_48":49,"_47":164,"_165":66,"_45":425,"_13":70,"_40":426,"_42":427,"_30":31,"_169":-5,"_255":428,"_33":429},{"_50":414,"_38":415,"_36":416,"_163":205,"_48":49,"_47":417,"_165":31,"_45":418,"_13":164,"_40":41,"_42":419,"_30":31,"_169":-5,"_255":420,"_33":421},{"_50":198,"_38":199,"_36":200,"_163":-5,"_48":49,"_47":164,"_165":66,"_45":411,"_13":70,"_40":191,"_42":-5,"_30":31,"_169":-5,"_255":412,"_33":413},{"_50":404,"_38":405,"_36":406,"_163":200,"_48":49,"_47":164,"_165":66,"_45":407,"_13":70,"_40":41,"_42":408,"_30":31,"_169":-5,"_255":409,"_33":410},{"_50":397,"_38":398,"_36":399,"_163":200,"_48":49,"_47":164,"_165":66,"_45":400,"_13":70,"_40":41,"_42":401,"_30":31,"_169":-5,"_255":402,"_33":403},{"_50":389,"_38":390,"_36":391,"_163":200,"_48":49,"_47":392,"_165":31,"_45":393,"_13":279,"_40":41,"_42":394,"_30":31,"_169":-5,"_255":395,"_33":396},{"_50":382,"_38":383,"_36":384,"_163":200,"_48":49,"_47":164,"_165":66,"_45":385,"_13":70,"_40":41,"_42":386,"_30":31,"_169":-5,"_255":387,"_33":388},{"_50":193,"_38":194,"_36":195,"_163":-5,"_48":49,"_47":164,"_165":66,"_45":379,"_13":70,"_40":191,"_42":-5,"_30":31,"_169":-5,"_255":380,"_33":381},{"_50":375,"_38":329,"_36":323,"_163":195,"_48":49,"_47":164,"_165":66,"_45":376,"_13":70,"_40":191,"_42":-5,"_30":31,"_169":-5,"_255":377,"_33":378},{"_50":366,"_38":367,"_36":368,"_163":323,"_48":49,"_47":369,"_165":31,"_45":370,"_13":371,"_40":41,"_42":372,"_30":31,"_169":-5,"_255":373,"_33":374},{"_50":359,"_38":360,"_36":361,"_163":323,"_48":49,"_47":164,"_165":66,"_45":362,"_13":70,"_40":41,"_42":363,"_30":31,"_169":-5,"_255":364,"_33":365},{"_50":352,"_38":353,"_36":354,"_163":323,"_48":49,"_47":164,"_165":66,"_45":355,"_13":70,"_40":41,"_42":356,"_30":31,"_169":-5,"_255":357,"_33":358},{"_50":345,"_38":346,"_36":347,"_163":323,"_48":49,"_47":164,"_165":66,"_45":348,"_13":70,"_40":41,"_42":349,"_30":31,"_169":-5,"_255":350,"_33":351},{"_50":337,"_38":338,"_36":339,"_163":323,"_48":49,"_47":340,"_165":31,"_45":341,"_13":279,"_40":41,"_42":342,"_30":31,"_169":-5,"_255":343,"_33":344},{"_50":330,"_38":331,"_36":332,"_163":323,"_48":49,"_47":164,"_165":66,"_45":333,"_13":70,"_40":41,"_42":334,"_30":31,"_169":-5,"_255":335,"_33":336},{"_50":320,"_38":321,"_36":322,"_163":323,"_48":49,"_47":324,"_165":66,"_45":325,"_13":70,"_40":41,"_42":326,"_30":31,"_169":-5,"_255":327,"_33":328},{"_50":313,"_38":314,"_36":315,"_163":195,"_48":49,"_47":164,"_165":66,"_45":316,"_13":70,"_40":26,"_42":317,"_30":31,"_169":-5,"_255":318,"_33":319},{"_50":306,"_38":307,"_36":308,"_163":195,"_48":49,"_47":164,"_165":66,"_45":309,"_13":70,"_40":41,"_42":310,"_30":31,"_169":-5,"_255":311,"_33":312},{"_50":299,"_38":300,"_36":301,"_163":195,"_48":49,"_47":164,"_165":66,"_45":302,"_13":70,"_40":41,"_42":303,"_30":31,"_169":-5,"_255":304,"_33":305},{"_50":292,"_38":293,"_36":294,"_163":195,"_48":49,"_47":164,"_165":66,"_45":295,"_13":70,"_40":41,"_42":296,"_30":31,"_169":-5,"_255":297,"_33":298},{"_50":284,"_38":285,"_36":286,"_163":195,"_48":49,"_47":287,"_165":66,"_45":288,"_13":164,"_40":41,"_42":289,"_30":31,"_169":-5,"_255":290,"_33":291},{"_50":276,"_38":277,"_36":278,"_163":195,"_48":49,"_47":279,"_165":66,"_45":280,"_13":164,"_40":41,"_42":281,"_30":31,"_169":-5,"_255":282,"_33":283},{"_50":187,"_38":188,"_36":189,"_163":-5,"_48":49,"_47":164,"_165":66,"_45":273,"_13":70,"_40":191,"_42":-5,"_30":31,"_169":-5,"_255":274,"_33":275},{"_50":264,"_38":265,"_36":266,"_163":189,"_48":49,"_47":267,"_165":31,"_45":268,"_13":269,"_40":41,"_42":270,"_30":31,"_169":-5,"_255":271,"_33":272},{"_50":180,"_38":181,"_36":182,"_163":-5,"_48":49,"_47":183,"_165":31,"_45":261,"_13":177,"_40":26,"_42":185,"_30":31,"_169":-5,"_255":262,"_33":263},{"_50":172,"_38":173,"_36":174,"_163":-5,"_48":49,"_47":175,"_165":31,"_45":258,"_13":177,"_40":26,"_42":178,"_30":31,"_169":-5,"_255":259,"_33":260},{"_50":161,"_38":162,"_36":64,"_163":-5,"_48":49,"_47":164,"_165":66,"_45":254,"_13":70,"_40":167,"_42":168,"_30":31,"_169":170,"_255":256,"_33":257},["D",1789082831202],"children",[],[162],["D",1789484890481],[],[173],["D",1789484890661],[],[181],"25f218dd-cb4f-4bea-8f58-371e8cec9f57","types.ts","01a08da5-899f-73ab-baf1-e4b182203ecc",10,["D",1789484573494],2,"export type BenchmarkStatus = \"running\" | \"completed\" | \"failed\";\nexport type MeasurementStatus = \"ok\" | \"error\";\n\nexport interface BenchmarkConfig {\n benchmarkVersion: string;\n label: string;\n modelId: string;\n runsPerProvider: number;\n maxOutputTokens: number;\n providerConcurrency: number;\n requestTimeoutMs: number;\n providers?: string[];\n}\n\n/**\n * \"provider\" rows pin one provider with `only: [name]`. \"strategy\" rows are\n * synthetic endpoints that exercise a way of calling the gateway (default\n * autorouting, `sort`, `order`, handpicked `only` sets…).\n */\nexport type EndpointKind = \"provider\" | \"strategy\";\n\nexport interface GatewayEndpoint {\n provider_name: string;\n kind?: EndpointKind;\n description?: string;\n gateway_options?: Record;\n name?: string;\n has_zdr?: boolean;\n has_no_training?: boolean;\n supports_implicit_caching?: boolean;\n pricing?: Record;\n latency_last_1h?: { p50?: number; p95?: number };\n throughput_last_1h?: { p50?: number; p95?: number };\n tags?: string[];\n [key: string]: unknown;\n}\n\nexport interface BenchmarkRunSummary {\n id: string;\n label: string;\n modelId: string;\n status: BenchmarkStatus;\n startedAt: string;\n completedAt: string | null;\n runsPerProvider: number;\n maxOutputTokens: number;\n providerCount: number;\n successfulMeasurements: number;\n failedMeasurements: number;\n error: string | null;\n}\n\nexport type StoredMeasurementOutcome = \"text\" | \"no_text\" | \"error\";\nexport type MeasurementErrorCode =\n | \"no_text\"\n | \"timeout\"\n | \"provider_error\"\n | null;\n\nexport interface BenchmarkMeasurement {\n runId: string;\n provider: string;\n attempt: number;\n status: MeasurementStatus;\n outcome: StoredMeasurementOutcome;\n errorCode: MeasurementErrorCode;\n startedAt: string;\n completedAt: string;\n ttftMs: number | null;\n durationMs: number | null;\n inputTokens: number | null;\n outputTokens: number | null;\n cacheReadTokens: number | null;\n cacheWriteTokens: number | null;\n costUsd: number | null;\n tokensPerSecond: number | null;\n responsePreview: string | null;\n error: string | null;\n providerMetadata: Record | null;\n providerInfo: GatewayEndpoint;\n}\n\nexport type MeasurementPublicStatus = \"success\" | \"error\";\n\nexport type MeasurementSortKey =\n | \"runStartedAt\"\n | \"provider\"\n | \"attempt\"\n | \"status\"\n | \"startedAt\"\n | \"ttftMs\"\n | \"durationMs\"\n | \"tokensPerSecond\"\n | \"inputTokens\"\n | \"outputTokens\"\n | \"cacheReadTokens\"\n | \"cacheWriteTokens\"\n | \"costUsd\"\n | \"message\";\n\nexport type MeasurementSortDirection = \"asc\" | \"desc\";\n\nexport interface MeasurementListItem {\n runId: string;\n runStartedAt: string;\n runStatus: BenchmarkStatus;\n provider: string;\n attempt: number;\n status: MeasurementPublicStatus;\n startedAt: string;\n completedAt: string;\n ttftMs: number | null;\n durationMs: number | null;\n inputTokens: number | null;\n outputTokens: number | null;\n cacheReadTokens: number | null;\n cacheWriteTokens: number | null;\n costUsd: number | null;\n tokensPerSecond: number | null;\n error: string | null;\n}\n\nexport interface MeasurementRunOption {\n id: string;\n startedAt: string;\n status: BenchmarkStatus;\n measurementCount: number;\n}\n\nexport interface MeasurementListResponse {\n items: MeasurementListItem[];\n total: number;\n limit: number;\n offset: number;\n providers: string[];\n attempts: number[];\n runs: MeasurementRunOption[];\n}\n\nexport interface BenchmarkRunDetail extends BenchmarkRunSummary {\n config: BenchmarkConfig;\n endpointSnapshot: GatewayEndpoint[];\n measurements: BenchmarkMeasurement[];\n}\n\nexport interface ProviderMetricPoint {\n runId: string;\n runStartedAt: string;\n successRate: number;\n attempts: number;\n successes: number;\n p50TtftMs: number | null;\n p95TtftMs: number | null;\n p50TokensPerSecond: number | null;\n p05TokensPerSecond: number | null;\n cacheHitRate: number | null;\n effectiveCostPerMillionTokensUsd: number | null;\n}\n\nexport interface ProviderResult {\n provider: string;\n kind: EndpointKind;\n /** Zero data retention per the Vercel endpoint catalogue (null = unknown). */\n hasZdr: boolean | null;\n /** Provider does not train on prompts per the catalogue (null = unknown). */\n hasNoTraining: boolean | null;\n /** providerOptions.gateway used by a strategy row; null for provider rows. */\n gatewayOptions: Record | null;\n /** Successful attempts by the provider the gateway finally routed to. */\n routedProviders: Record;\n /** Successful attempts where the gateway retried after a provider failure. */\n recoveredAttempts: number;\n runCount: number;\n attempts: number;\n successes: number;\n failures: number;\n successRate: number;\n p50TtftMs: number | null;\n p95TtftMs: number | null;\n p50DurationMs: number | null;\n p95DurationMs: number | null;\n p50TokensPerSecond: number | null;\n p05TokensPerSecond: number | null;\n cacheEligibleAttempts: number;\n cacheHits: number;\n cacheHitRate: number | null;\n effectiveCostPerMillionTokensUsd: number | null;\n history: ProviderMetricPoint[];\n}\n\nexport interface BenchmarkResults {\n modelId: string;\n runCount: number;\n attemptCount: number;\n successfulAttempts: number;\n failedAttempts: number;\n firstRunAt: string | null;\n lastCompletedAt: string | null;\n providers: ProviderResult[];\n}\n",[],[188,265],["D",1789082831202],[250],[188],"70e5d894-3ea5-4a94-9391-a86dad9e083f","root.tsx","01a08da5-899f-73ab-baf1-df2fd8ed278d",6,["D",1789082872323],"/** @jsxImportSource npm:hono@4/jsx */\nimport { raw } from \"npm:hono@4/html\";\nimport { immutableFileUrl } from \"https://esm.town/v/std/utils/index.ts\";\n\nexport function Root() {\n return (\n <>\n {raw(\"\")}\n \n \n \n \n \n \n DeepSeek V4.1 Flash Provider Benchmark\n

Diff · v41