{"id":"preference-dpo-sft-trace-distillation","slug":"preference-dpo-sft-trace-distillation","name":"Trajectory SFT, DPO & GRPO Distillation (Teacher-to-Student Agent Fine-Tuning)","shortName":"Trajectory SFT / DPO / GRPO distillation","problemStatement":"Distill verified high-scoring trajectories from an expensive frontier teacher model (plus rejected failure pairs) into a smaller, faster open-weight student model via LoRA SFT, Direct Preference Optimization (DPO), or Verifier-Guided GRPO.","whenNotTopPick":"Ultra-low-latency serverless LoRA multi-tenant deployment after tuning — Fireworks AI; experiment tracking, dataset lineage, and reward-curve telemetry across RL/DPO sweeps — Weights & Biases; raw multi-node H100/B200 clusters for custom Megatron/vLLM RL loops — Lambda or Modal.","categoryIds":["cat-inference-apis","cat-gpu-cloud"],"depotPickProviderId":"together","depotPickRationale":"Together AI offers turnkey serverless and dedicated fine-tuning (LoRA SFT, DPO, and continued pretraining across Qwen, Llama, and DeepSeek families) with direct one-click deployment onto Together serverless endpoints, avoiding idle GPU burn while serving distilled agent adapters.","runnersUpIds":["fireworks","weights-biases","modal"],"weights":{"capabilityFit":0.28,"pricePerformance":0.2,"latencyReliability":0.14,"developerDx":0.16,"privacyPosture":0.1,"lockInRisk":0.06,"maturity":0.06},"methodologyNotes":"Distillation only helps when training pairs are filtered by a deterministic verifier (unit tests, schema validation, or Countersign receipt). Training on unverified teacher traces amplifies hallucinations. Always benchmark the distilled student against the teacher on `costPerSolvedTaskUsd` and `pass@1`.","tags":["learning","fine-tuning","distillation","dpo","grpo","sft","lora","self-improvement"],"scoreMargin":-0.15,"isNearTie":true,"pickConfidenceBand":"near_tie_co_pick","nearTie":{"isNearTie":true,"scoreMargin":-0.15,"threshold":0.35,"pickConfidenceBand":"near_tie_co_pick","coPickProviderId":"fireworks","coPickProviderName":"Fireworks AI","coPickScore":8.4,"decidingFactor":"pricePerformance (9 vs 8.8)","methodologyWarning":"Top pick (8.25) and #2 runner-up (8.4) are separated by 0.15 points (<= 0.35 weight-sensitivity band). Treat them as co-picks and use the workload tie-breaker rule below rather than the raw composite decimal.","tieBreakerRule":"Ultra-low-latency serverless LoRA multi-tenant deployment after tuning — Fireworks AI; experiment tracking, dataset lineage, and reward-curve telemetry across RL/DPO sweeps — Weights & Biases; raw multi-node H100/B200 clusters for custom Megatron/vLLM RL loops — Lambda or Modal."},"evidenceStatus":{"decisionRole":"independent_prior_and_candidate_generator","positioning":"The independent, agent-readable prior for choosing AI infrastructure — not a universal router, generic marketplace, or synthetic leaderboard.","freshnessThresholdDays":30,"freshness_policy":{"policy_version":"agentdepot.freshness_policy.v1","threshold_days":30},"freshness_state":"fresh","editorial_freshness_state":"stale_prior","unqualified_current_pick_allowed":true,"pick_qualification":"qualified_current_pick","promotionClass":"deterministic_source_reverified","evidenceClass":"editorial_prior","staleDowngraded":false,"effectiveConfidence":"medium_reverified","confidenceHalfWidth":0.5,"pickScoreAgeDays":25,"editorialJudgmentAgeDays":44,"pricingAgeDays":25,"isPickScoreStale":false,"isPricingStale":false,"editorialJudgmentAsOf":"2026-09-14","pricingVerifiedAt":"2026-10-03","outcomeVerifiedAt":null,"effectiveVerifiedAsOf":"2026-10-03","downgradeReasons":[],"stalenessWarning":"Editorial judgment scored on 2026-09-14 (44d ago); promoted to deterministic_source_reverified on 2026-10-03 via deterministic pricing/source gate.","promotionGate":{"deterministicPricingCheck":true,"approvedScoreReaffirm":false,"pendingHumanReviewBlock":false,"zeroUnverifiedSkus":true,"costEfficiencyConsistent":true,"lockInAndReliabilityConsistent":true,"secondModelOrRuleAudit":true,"countersignReceiptIds":[],"otsAttestationHead":"b32634d5784cbfe663be0a16cce4f60a3ff1c3d4490437be00361cb4ae5d5d59","sourceCheckDate":"2026-10-27","activePricingRowsCount":10,"freeResourcesCount":0,"runnerUpPricingRowsCount":0},"claimProvenance":{"claim_id":"claim:preference-dpo-sft-trace-distillation:together:depot-pick","provider_id":"together","jtbd_id":"preference-dpo-sft-trace-distillation","primary_url":"https://www.together.ai/pricing","source_type":"primary_pricing","source_revision":"sha256:ef94b3764157f3e8","observed_at":"2026-10-28","checked_at":"2026-10-28","claim_support_state":"claim_specific_supported","max_allowed_evidence_class":"replicated_outcome","defect_codes":[],"sources":[{"original_url":"https://www.together.ai","final_url":"https://www.together.ai","redirected":false,"http_status":200,"reachable":true,"source_type":"generic_homepage","is_claim_specific":false,"source_revision":"sha256:649d028760d3f2fc","observed_at":"2026-10-28","checked_at":"2026-10-28"},{"original_url":"https://www.together.ai/pricing","final_url":"https://www.together.ai/pricing","redirected":false,"http_status":200,"reachable":true,"source_type":"primary_pricing","is_claim_specific":true,"source_revision":"sha256:ef94b3764157f3e8","observed_at":"2026-10-28","checked_at":"2026-10-28"},{"original_url":"https://docs.together.ai","final_url":"https://docs.together.ai","redirected":false,"http_status":200,"reachable":true,"source_type":"generic_homepage","is_claim_specific":false,"source_revision":"sha256:1f1159576bb599e4","observed_at":"2026-10-28","checked_at":"2026-10-28"}]},"verifiedOutcomeSummary":{"totalVerifiedReceipts":0,"byEvidenceTier":{"E0_ASSERTION":0,"E1_SOURCE_BOUND":0,"E2_DELIVERY":0,"E3_INDEPENDENT_OUTCOME":0,"E4_REPLICATED_OUTCOME":0},"bySubjectLayer":{"provider":0,"product":0,"model_version":0,"runtime_route":0}}},"pickScoreDetails":{"scoreRange":[7.75,8.75],"formattedScore":"8.3 ± 0.5 (7.8–8.8 · editorial_prior)","evidenceClass":"editorial_prior","promotionClass":"deterministic_source_reverified","staleDowngraded":false,"freshState":"fresh","freshness_state":"fresh","effectiveConfidence":"medium_reverified","confidenceInterval":{"low":7.75,"mid":8.25,"high":8.75,"halfWidth":0.5,"confidenceLevel":0.8,"basis":"deterministic_source_reverified_interval_0.5"},"editorialJudgmentAsOf":"2026-10-28","sourceAgeDays":44,"pricingVerifiedAt":"2026-10-03","pricingAgeDays":25,"outcomeVerifiedAt":null,"effectiveVerifiedAsOf":"2026-10-03","dissentingEvidence":["Billing tiers/spend gates can surprise mid-build","Quality variance across catalog models","Support uneven at small spend","When not to use top pick: Ultra-low-latency serverless LoRA multi-tenant deployment after tuning — Fireworks AI; experiment tracking, dataset lineage, and reward-curve telemetry across RL/DPO sweeps — Weights & Biases; raw multi-node H100/B200 clusters for custom Megatron/vLLM RL loops — Lambda or Modal.","Runner-up Fireworks AI (fireworks) is the primary alternative when constraints disqualify Together AI."],"whatWouldChangePick":["Caller constraint matching: Ultra-low-latency serverless LoRA multi-tenant deployment after tuning — Fireworks AI; experiment tracking, dataset lineage, and reward-curve telemetry across RL/DPO sweeps — Weights & Biases; raw multi-node H100/B200 clusters for custom Megatron/vLLM RL loops — Lambda or Modal.","Live quota or rate-limit exhaustion on together (pass ?exhausted=together or liveQuota.exhaustedProviders=[\"together\"] to promote fireworks).","Active subscription or enterprise commit on fireworks combined with low switching-cost tolerance (existingSubscriptions + switchingCost.maxMigrationEffort=\"low\").","Sourced price increase >20% on https://www.together.ai or unverified SKU drift detected by npm run check:sources."],"claimProvenance":{"claim_id":"claim:preference-dpo-sft-trace-distillation:together:depot-pick","provider_id":"together","jtbd_id":"preference-dpo-sft-trace-distillation","primary_url":"https://www.together.ai/pricing","source_type":"primary_pricing","source_revision":"sha256:ef94b3764157f3e8","observed_at":"2026-10-28","checked_at":"2026-10-28","claim_support_state":"claim_specific_supported","max_allowed_evidence_class":"replicated_outcome","defect_codes":[],"sources":[{"original_url":"https://www.together.ai","final_url":"https://www.together.ai","redirected":false,"http_status":200,"reachable":true,"source_type":"generic_homepage","is_claim_specific":false,"source_revision":"sha256:649d028760d3f2fc","observed_at":"2026-10-28","checked_at":"2026-10-28"},{"original_url":"https://www.together.ai/pricing","final_url":"https://www.together.ai/pricing","redirected":false,"http_status":200,"reachable":true,"source_type":"primary_pricing","is_claim_specific":true,"source_revision":"sha256:ef94b3764157f3e8","observed_at":"2026-10-28","checked_at":"2026-10-28"},{"original_url":"https://docs.together.ai","final_url":"https://docs.together.ai","redirected":false,"http_status":200,"reachable":true,"source_type":"generic_homepage","is_claim_specific":false,"source_revision":"sha256:1f1159576bb599e4","observed_at":"2026-10-28","checked_at":"2026-10-28"}]},"calibratedScore":{"point":8.25,"band":[7.75,8.75],"interval":{"lower":7.75,"upper":8.75,"half_width":0.5},"score_kind":"source_bound_prior_band","evidence_class":"source_bound","confidence":"low","confidence_ceiling":"medium","age_days":25,"dissent":["Billing tiers/spend gates can surprise mid-build","Quality variance across catalog models","Support uneven at small spend","When not to use top pick: Ultra-low-latency serverless LoRA multi-tenant deployment after tuning — Fireworks AI; experiment tracking, dataset lineage, and reward-curve telemetry across RL/DPO sweeps — Weights & Biases; raw multi-node H100/B200 clusters for custom Megatron/vLLM RL loops — Lambda or Modal.","Runner-up Fireworks AI (fireworks) is the primary alternative when constraints disqualify Together AI."],"what_would_change_this":["Caller constraint matching: Ultra-low-latency serverless LoRA multi-tenant deployment after tuning — Fireworks AI; experiment tracking, dataset lineage, and reward-curve telemetry across RL/DPO sweeps — Weights & Biases; raw multi-node H100/B200 clusters for custom Megatron/vLLM RL loops — Lambda or Modal.","Live quota or rate-limit exhaustion on together (pass ?exhausted=together or liveQuota.exhaustedProviders=[\"together\"] to promote fireworks).","Active subscription or enterprise commit on fireworks combined with low switching-cost tolerance (existingSubscriptions + switchingCost.maxMigrationEffort=\"low\").","Sourced price increase >20% on https://www.together.ai or unverified SKU drift detected by npm run check:sources."],"subject_layer":"provider"},"subject_ref":{"layer":"provider","provider_id":"together","product_id":"together-core","model_id":null,"version":null,"route_class":null,"inheritance_policy":"no_cross_layer_inheritance"}},"runtimeRoutes":[{"routeId":"preference-dpo-sft-trace-distillation:together:primary","roleForJob":"depot_pick","provider":{"id":"together","slug":"together","name":"Together AI","website":"https://www.together.ai","status":"active"},"model":{"representativeModelId":"DeepSeek V4 Pro 0813","skuId":"together-deepseek-v4-pro-0813-input","unit":"usd_per_million_input_tokens","amountUsd":1.32,"sourceUrl":"https://www.together.ai/pricing","asOf":"2026-10-28"},"servingRoute":{"primaryRouteType":"direct_first_party_api","endpointPattern":"https://www.together.ai (direct first party api)","aggregatorAlternatives":[{"aggregatorId":"direct-only","name":"Together AI Direct API","routeType":"first_party","tradeoff":"Direct vendor SDK/REST contract; pair with runner-up for multi-vendor failover."}]},"accountAndQuota":{"freeTierAvailable":false,"freeTierSummary":null,"rateLimitsSummary":"See /api/limits?provider=together","subscriptionFirstEligible":true,"documentedPlansCount":0},"regions":{"supportedRegions":["us","global"],"multiRegionFailover":true,"euResidencyAvailable":true},"toolSchemaSupport":{"strictJsonSchema":true,"parallelToolCalls":true,"mcpServerAvailable":false,"streamingSupported":true},"latencyTier":{"tier":"standard_1s_to_5s","latencyReliabilityScore":8,"coldStartNotes":"Standard warm pool; check provider SLA for burst concurrency"},"privacyConfig":{"privacyPostureScore":7.2,"zeroDataRetentionAvailable":false,"trainsOnApiDataByDefault":false,"hipaaEligible":false,"selfHostable":false},"availabilityStatus":{"providerStatus":"active","verifiedActiveSkus":10,"unverifiedSkus":2,"liveFailoverTargetId":"fireworks"},"subject_ref":{"layer":"runtime_route","subject_id":"runtime_route:together:DeepSeek V4 Pro 0813:direct_first_party_api:us","provider_id":"together","product_id":"together-core","model_id":"DeepSeek V4 Pro 0813","version":"2026-10-28","route_class":"direct_first_party_api","region":"us","score_status":"layer_evidenced","layer_score":"unknown","inherited_from_parent_layer":false,"inheritance_policy":"no_cross_layer_inheritance"},"distinctRouteSubjects":[{"layer":"runtime_route","subject_id":"runtime_route:together:DeepSeek V4 Pro 0813:direct_api:us","provider_id":"together","product_id":"together-core","model_id":"DeepSeek V4 Pro 0813","version":"2026-10-28","route_class":"direct_api","region":"us","score_status":"layer_evidenced","layer_score":"unknown","inherited_from_parent_layer":false,"inheritance_policy":"no_cross_layer_inheritance"}]},{"routeId":"preference-dpo-sft-trace-distillation:fireworks:primary","roleForJob":"runner_up","provider":{"id":"fireworks","slug":"fireworks","name":"Fireworks AI","website":"https://fireworks.ai","status":"active"},"model":{"representativeModelId":"Embeddings up to 150M params","skuId":"fireworks-embed-to-150m","unit":"usd_per_million_input_tokens","amountUsd":0.008,"sourceUrl":"https://fireworks.ai/pricing","asOf":"2026-10-28"},"servingRoute":{"primaryRouteType":"direct_first_party_api","endpointPattern":"https://fireworks.ai (direct first party api)","aggregatorAlternatives":[{"aggregatorId":"openrouter","name":"OpenRouter","routeType":"unified_openai_proxy","tradeoff":"Automatic failover to Together/DeepInfra when rate-limited"}]},"accountAndQuota":{"freeTierAvailable":false,"freeTierSummary":null,"rateLimitsSummary":"See /api/limits?provider=fireworks","subscriptionFirstEligible":true,"documentedPlansCount":0},"regions":{"supportedRegions":["us","global"],"multiRegionFailover":true,"euResidencyAvailable":true},"toolSchemaSupport":{"strictJsonSchema":true,"parallelToolCalls":true,"mcpServerAvailable":true,"streamingSupported":true},"latencyTier":{"tier":"standard_1s_to_5s","latencyReliabilityScore":8.4,"coldStartNotes":"Standard warm pool; check provider SLA for burst concurrency"},"privacyConfig":{"privacyPostureScore":7.5,"zeroDataRetentionAvailable":false,"trainsOnApiDataByDefault":false,"hipaaEligible":false,"selfHostable":false},"availabilityStatus":{"providerStatus":"active","verifiedActiveSkus":14,"unverifiedSkus":0,"liveFailoverTargetId":"together"},"subject_ref":{"layer":"runtime_route","subject_id":"runtime_route:fireworks:Embeddings up to 150M params:direct_first_party_api:us","provider_id":"fireworks","product_id":"fireworks-core","model_id":"Embeddings up to 150M params","version":"2026-10-28","route_class":"direct_first_party_api","region":"us","score_status":"layer_evidenced","layer_score":"unknown","inherited_from_parent_layer":false,"inheritance_policy":"no_cross_layer_inheritance"},"distinctRouteSubjects":[{"layer":"runtime_route","subject_id":"runtime_route:fireworks:Embeddings up to 150M params:direct_api:us","provider_id":"fireworks","product_id":"fireworks-core","model_id":"Embeddings up to 150M params","version":"2026-10-28","route_class":"direct_api","region":"us","score_status":"layer_evidenced","layer_score":"unknown","inherited_from_parent_layer":false,"inheritance_policy":"no_cross_layer_inheritance"},{"layer":"runtime_route","subject_id":"runtime_route:fireworks:Embeddings up to 150M params:openrouter:global","provider_id":"fireworks","product_id":"fireworks-via-openrouter","model_id":"Embeddings up to 150M params","version":"2026-10-28","route_class":"openrouter","region":"global","score_status":"layer_evidenced","layer_score":"unknown","inherited_from_parent_layer":false,"inheritance_policy":"no_cross_layer_inheritance"}]},{"routeId":"preference-dpo-sft-trace-distillation:weights-biases:primary","roleForJob":"runner_up","provider":{"id":"weights-biases","slug":"weights-biases","name":"Weights & Biases","website":"https://wandb.ai","status":"active"},"model":{"representativeModelId":"Weights & Biases Primary Runtime","skuId":null,"unit":null,"amountUsd":null,"sourceUrl":"https://wandb.ai","asOf":"2026-10-28"},"servingRoute":{"primaryRouteType":"managed_cloud_service","endpointPattern":"https://wandb.ai (managed cloud service)","aggregatorAlternatives":[{"aggregatorId":"direct-only","name":"Weights & Biases Direct API","routeType":"first_party","tradeoff":"Direct vendor SDK/REST contract; pair with runner-up for multi-vendor failover."}]},"accountAndQuota":{"freeTierAvailable":false,"freeTierSummary":null,"rateLimitsSummary":"See /api/limits?provider=weights-biases","subscriptionFirstEligible":true,"documentedPlansCount":0},"regions":{"supportedRegions":["global"],"multiRegionFailover":true,"euResidencyAvailable":true},"toolSchemaSupport":{"strictJsonSchema":true,"parallelToolCalls":false,"mcpServerAvailable":false,"streamingSupported":true},"latencyTier":{"tier":"interactive_sub_1s","latencyReliabilityScore":9,"coldStartNotes":"Standard warm pool; check provider SLA for burst concurrency"},"privacyConfig":{"privacyPostureScore":7.8,"zeroDataRetentionAvailable":false,"trainsOnApiDataByDefault":false,"hipaaEligible":false,"selfHostable":false},"availabilityStatus":{"providerStatus":"active","verifiedActiveSkus":0,"unverifiedSkus":0,"liveFailoverTargetId":"together"},"subject_ref":{"layer":"runtime_route","subject_id":"runtime_route:weights-biases:weights-biases-primary:managed_cloud_service:global","provider_id":"weights-biases","product_id":"weights-biases-core","model_id":"weights-biases-primary","version":"2026-10-28","route_class":"managed_cloud_service","region":"global","score_status":"layer_evidenced","layer_score":"unknown","inherited_from_parent_layer":false,"inheritance_policy":"no_cross_layer_inheritance"},"distinctRouteSubjects":[{"layer":"runtime_route","subject_id":"runtime_route:weights-biases:weights-biases-primary:direct_api:global","provider_id":"weights-biases","product_id":"weights-biases-core","model_id":"weights-biases-primary","version":"2026-10-28","route_class":"direct_api","region":"global","score_status":"layer_evidenced","layer_score":"unknown","inherited_from_parent_layer":false,"inheritance_policy":"no_cross_layer_inheritance"}]},{"routeId":"preference-dpo-sft-trace-distillation:modal:primary","roleForJob":"runner_up","provider":{"id":"modal","slug":"modal","name":"Modal","website":"https://modal.com","status":"active"},"model":{"representativeModelId":"Nvidia B300","skuId":"modal-nvidia-b300-usd-per-gpu-second","unit":"usd_per_gpu_second","amountUsd":0.001972,"sourceUrl":"https://modal.com/pricing","asOf":"2026-10-27"},"servingRoute":{"primaryRouteType":"managed_cloud_service","endpointPattern":"https://modal.com (managed cloud service)","aggregatorAlternatives":[{"aggregatorId":"direct-only","name":"Modal Direct API","routeType":"first_party","tradeoff":"Direct vendor SDK/REST contract; pair with runner-up for multi-vendor failover."}]},"accountAndQuota":{"freeTierAvailable":false,"freeTierSummary":null,"rateLimitsSummary":"See /api/limits?provider=modal","subscriptionFirstEligible":true,"documentedPlansCount":0},"regions":{"supportedRegions":["us","global"],"multiRegionFailover":true,"euResidencyAvailable":true},"toolSchemaSupport":{"strictJsonSchema":true,"parallelToolCalls":false,"mcpServerAvailable":true,"streamingSupported":true},"latencyTier":{"tier":"standard_1s_to_5s","latencyReliabilityScore":8.6,"coldStartNotes":"Standard warm pool; check provider SLA for burst concurrency"},"privacyConfig":{"privacyPostureScore":7.8,"zeroDataRetentionAvailable":false,"trainsOnApiDataByDefault":false,"hipaaEligible":false,"selfHostable":false},"availabilityStatus":{"providerStatus":"active","verifiedActiveSkus":18,"unverifiedSkus":0,"liveFailoverTargetId":"together"},"subject_ref":{"layer":"runtime_route","subject_id":"runtime_route:modal:Nvidia B300:managed_cloud_service:us","provider_id":"modal","product_id":"modal-core","model_id":"Nvidia B300","version":"2026-10-27","route_class":"managed_cloud_service","region":"us","score_status":"layer_evidenced","layer_score":"unknown","inherited_from_parent_layer":false,"inheritance_policy":"no_cross_layer_inheritance"},"distinctRouteSubjects":[{"layer":"runtime_route","subject_id":"runtime_route:modal:Nvidia B300:direct_api:us","provider_id":"modal","product_id":"modal-core","model_id":"Nvidia B300","version":"2026-10-27","route_class":"direct_api","region":"us","score_status":"layer_evidenced","layer_score":"unknown","inherited_from_parent_layer":false,"inheritance_policy":"no_cross_layer_inheritance"}]}],"depth":{"whenNotToUse":{"text":"Ultra-low-latency serverless LoRA multi-tenant deployment after tuning — Fireworks AI; experiment tracking, dataset lineage, and reward-curve telemetry across RL/DPO sweeps — Weights & Biases; raw multi-node H100/B200 clusters for custom Megatron/vLLM RL loops — Lambda or Modal.","source":"editorial `whenNotTopPick` (data/jtbds.json)"},"whenNotToUseReason":null,"runnerUp":{"providerId":"fireworks","name":"Fireworks AI","overallForJob":8.4,"pickOverallForJob":8.25,"beatsPickOn":["capabilityFit","latencyReliability","developerDx","privacyPosture"],"source":"first resolvable entry of the job's editorially ranked `runnersUpIds`; scores are stored job-weighted scores"},"runnerUpReason":null,"lockIn":{"pickLockInRisk":3.5,"runnerUpLockInRisk":3.5,"scale":"0–10 stored curated estimate; higher = more lock-in (worse)","pickExitSignals":["serves open-weight models (provider tag `open-models`)"],"runnerUpExitSignals":["serves open-weight models (provider tag `open-models`)"],"pickNotes":[],"selfHostFallback":[],"migration":"Pick exit signals: serves open-weight models (provider tag `open-models`).","migrationReason":null,"derivedFrom":"stored lockInRisk scores, provider tags (oss, self-host, openai-compat, open-models), free rows that list this job, and verbatim provider notes. No migration guide was invented."}},"freshness":{"pick":{"score":{"asOf":"2026-10-28","ageDays":1,"stale":false,"staleAfterDays":30,"reason":null},"lastReviewed":{"asOf":"2026-10-28","ageDays":1,"stale":false,"staleAfterDays":30,"reason":null},"sourceCheck":{"checkedAt":"2026-10-28","sourceUrls":3,"sourcesReachable":3,"pricingRows":12,"pricingRowsReconfirmed":8,"note":"A source re-check confirms the page is reachable and, for pricing rows, that the stored amount and SKU name appear on it. It does not re-derive or change the score.","reason":null}}},"whenNotTopPickRules":[{"ifConstraint":"latency:ultra_low_ttft","recommendedProviderId":"fireworks","activeModelOrSku":"accounts/fireworks/models/glm-5p3-flash","reason":"Ultra-low-latency serverless LoRA multi-tenant deployment after tuning — Fireworks AI"},{"ifConstraint":"exit:self_host_or_open_weights","recommendedProviderId":"weights-biases","activeModelOrSku":"wandb-teams","reason":"experiment tracking, dataset lineage, and reward-curve telemetry across RL/DPO sweeps — Weights & Biases"},{"ifConstraint":"escape_hatch_clause_3","recommendedProviderId":"modal","activeModelOrSku":"modal-h100-gpu-second","reason":"raw multi-node H100/B200 clusters for custom Megatron/vLLM RL loops — Lambda or Modal."}],"bootstrap":{"authType":"api_key","envVars":["TOGETHER_API_KEY"],"signupUrl":"https://api.together.xyz/signin","apiKeyUrl":"https://api.together.xyz/settings/api-keys","docsUrl":"https://docs.together.ai/docs/quickstart","mcpServer":null,"cliInit":null,"sdkInstall":{"npm":"npm install together-ai","pip":"pip install together"},"installCommand":"npm install together-ai","activeModelOrSku":"deepseek-ai/DeepSeek-V4.1-Flash","canonicalSnippet":"import Together from \"together-ai\";\nconst together = new Together({ apiKey: process.env.TOGETHER_API_KEY });","verifyCommand":"curl -sS https://api.together.xyz/v1/models -H \"Authorization: Bearer $TOGETHER_API_KEY\""},"quickstartSpec":{"installCommand":"npm install together-ai","envVars":["TOGETHER_API_KEY"],"activeModelOrSku":"deepseek-ai/DeepSeek-V4.1-Flash","canonicalSnippet":"import Together from \"together-ai\";\nconst together = new Together({ apiKey: process.env.TOGETHER_API_KEY });","sdkInstall":{"npm":"npm install together-ai","pip":"pip install together"},"verifyCommand":"curl -sS https://api.together.xyz/v1/models -H \"Authorization: Bearer $TOGETHER_API_KEY\"","apiKeyUrl":"https://api.together.xyz/settings/api-keys","docsUrl":"https://docs.together.ai/docs/quickstart","mcpServer":null,"bootstrapReason":null},"operationalGotchas":[{"code":"hidden_spend_tier_gates","severity":"warning","summary":"Billing tiers/spend gates can surprise mid-build","sourceDate":"2026-10-28"},{"code":"sku_rotation_risk","severity":"warning","summary":"Quality variance across catalog models","sourceDate":"2026-10-28"}],"relatedPacks":[]}