Files
waggle-os/litellm-config.yaml
shumeiko.it 16eaf01e14 chore: scrub personal names, maintainer URLs and local paths (P2.2)
Sample data in tests, prompts and UI placeholders uses Alex Morgan
instead of the maintainer's name; decision attributions say "owner".
Personal details in two test fixtures are replaced. Local paths
(D:/Projects/...) and paths into the maintainer's planning repository
are reworded. Clone and install URLs in CONTRIBUTING.md and the getting
started guide use placeholders until the public repository location is
decided (Q9). The Benchmarks page no longer names the maintainer's
hive-mind repository, and the upstream fixes log names the main Waggle
repository instead of an internal git host.
2026-11-01 00:50:13 +01:00

499 lines
22 KiB
YAML

model_list:
# ── Anthropic ──────────────────────────────────────────────────
# Sprint 10 Task 1.2 (2026-04-21): Sonnet 4.6 dated-snapshot suffix
# `-20250514` was never a valid Claude API ID for the 4.6 family (docs
# list plain `claude-sonnet-4-6` as current ID + alias as of 2026-04-21
# per https://platform.claude.com/docs/en/docs/about-claude/models/overview).
# Repointed to the plain alias. Migration note: ops/litellm/README.md.
# refs sprint-10-brief §1.1
- model_name: claude-sonnet-4-6
litellm_params:
model: anthropic/claude-sonnet-4-6
api_key: os.environ/ANTHROPIC_API_KEY
- model_name: claude-opus-4-6
litellm_params:
model: anthropic/claude-opus-4-6
api_key: os.environ/ANTHROPIC_API_KEY
- model_name: claude-haiku-4-5
litellm_params:
model: anthropic/claude-haiku-4-5-20251001
api_key: os.environ/ANTHROPIC_API_KEY
# Legacy aliases (backwards compat). Both now point to the plain
# `claude-sonnet-4-6` alias per Sprint 10 Task 1.2 (see note above).
- model_name: claude-3-5-sonnet-20241022
litellm_params:
model: anthropic/claude-sonnet-4-6
api_key: os.environ/ANTHROPIC_API_KEY
- model_name: anthropic/claude-sonnet-4-6
litellm_params:
model: anthropic/claude-sonnet-4-6
api_key: os.environ/ANTHROPIC_API_KEY
- model_name: anthropic/claude-opus-4.6
litellm_params:
model: anthropic/claude-opus-4-6
api_key: os.environ/ANTHROPIC_API_KEY
# ── OpenAI ─────────────────────────────────────────────────────
- model_name: gpt-4o
litellm_params:
model: openai/gpt-4o
api_key: os.environ/OPENAI_API_KEY
- model_name: gpt-4o-mini
litellm_params:
model: openai/gpt-4o-mini
api_key: os.environ/OPENAI_API_KEY
- model_name: o3
litellm_params:
model: openai/o3
api_key: os.environ/OPENAI_API_KEY
# ── Google Gemini ──────────────────────────────────────────────
# Use stable slug (not dated preview). Google deprecated -preview-05-20
# and -preview-06-05 — verified 2026-04-21 via live /v1beta/models list.
- model_name: gemini-2.5-pro
litellm_params:
model: gemini/gemini-2.5-pro
api_key: os.environ/GEMINI_API_KEY
- model_name: gemini-2.5-flash
litellm_params:
model: gemini/gemini-2.5-flash
api_key: os.environ/GEMINI_API_KEY
# ── xAI (Grok) ────────────────────────────────────────────────
- model_name: grok-3
litellm_params:
model: xai/grok-3
api_key: os.environ/XAI_API_KEY
- model_name: grok-3-mini
litellm_params:
model: xai/grok-3-mini
api_key: os.environ/XAI_API_KEY
# ── DeepSeek ───────────────────────────────────────────────────
- model_name: deepseek-chat
litellm_params:
model: deepseek/deepseek-chat
api_key: os.environ/DEEPSEEK_API_KEY
- model_name: deepseek-reasoner
litellm_params:
model: deepseek/deepseek-reasoner
api_key: os.environ/DEEPSEEK_API_KEY
# ── Perplexity ─────────────────────────────────────────────────
- model_name: perplexity-sonar
litellm_params:
model: perplexity/sonar
api_key: os.environ/PERPLEXITY_API_KEY
- model_name: perplexity-sonar-pro
litellm_params:
model: perplexity/sonar-pro
api_key: os.environ/PERPLEXITY_API_KEY
# ── OpenRouter (fallback to any model) ─────────────────────────
- model_name: openrouter/auto
litellm_params:
model: openrouter/auto
api_key: os.environ/OPENROUTER_API_KEY
# OpenRouter-routed aliases — usable when native provider key is missing
- model_name: claude-opus-4-7-via-openrouter
litellm_params:
model: openrouter/anthropic/claude-opus-4.7
api_key: os.environ/OPENROUTER_API_KEY
- model_name: claude-sonnet-4-6-via-openrouter
litellm_params:
model: openrouter/anthropic/claude-sonnet-4.6
api_key: os.environ/OPENROUTER_API_KEY
# ── Moonshot / Kimi ────────────────────────────────────────────
# Legacy endpoint (kept for backwards compatibility)
- model_name: moonshot-v1
litellm_params:
model: moonshot/moonshot-v1-128k
api_key: os.environ/MOONSHOT_API_KEY
# Current Kimi K2 family — routed via OpenRouter because the slugs
# moonshotai/kimi-k2.5, kimi-k2-thinking, kimi-k2.6 are all live on
# openrouter.ai/api/v1/models (verified 2026-04-21). Native
# platform.moonshot.cn slugs differ (kimi-k2-0905 etc) and the API
# surface drifts; OR normalizes.
- model_name: kimi-k2
litellm_params:
model: openrouter/moonshotai/kimi-k2
api_key: os.environ/OPENROUTER_API_KEY
- model_name: kimi-k2.5
litellm_params:
model: openrouter/moonshotai/kimi-k2.5
api_key: os.environ/OPENROUTER_API_KEY
# Reasoning variant (chain-of-thought exposed). Keep <think>...</think>
# tags in message history — see the MiniMax M2.7 note below.
- model_name: kimi-k2-thinking
litellm_params:
model: openrouter/moonshotai/kimi-k2-thinking
api_key: os.environ/OPENROUTER_API_KEY
# Kimi K2.6 — Moonshot's latest (added 2026-04-21). OR slug verified.
- model_name: kimi-k2.6
litellm_params:
model: openrouter/moonshotai/kimi-k2.6
api_key: os.environ/OPENROUTER_API_KEY
# Native Moonshot fallback (used if OR is rate-limited). The user's
# MOONSHOT_API_KEY is registered on the intl tenant (api.moonshot.ai,
# NOT api.moonshot.cn — the .cn endpoint returned empty on probe
# 2026-04-21). Using `moonshot-v1-auto` instead of pinning to a k2.x
# variant: kimi-k2.6 enforces temperature=1 at the API layer, which
# breaks generic smoke tests and deterministic eval harnesses. The
# `-auto` slug lets Moonshot pick the right variant and accepts
# standard temperature ranges.
- model_name: kimi-k2-native
litellm_params:
model: openai/moonshot-v1-auto
api_base: https://api.moonshot.ai/v1
api_key: os.environ/MOONSHOT_API_KEY
# UI aliases — providers.ts surfaces `kimi-2.5` and `kimi-2.5-thinking`
# in the Waggle UI model picker. Route them to the live OR entries so
# the UI doesn't 404.
- model_name: kimi-2.5
litellm_params:
model: openrouter/moonshotai/kimi-k2.5
api_key: os.environ/OPENROUTER_API_KEY
- model_name: kimi-2.5-thinking
litellm_params:
model: openrouter/moonshotai/kimi-k2-thinking
api_key: os.environ/OPENROUTER_API_KEY
# ── Alibaba / Qwen ────────────────────────────────────────────
# Sprint 10 Task 1.4 (2026-04-21): classic DashScope Model Studio API
# key (sk-…) provisioned on the intl tenant. Smoke-verified: target
# `qwen3.6-35b-a3b` returns HTTP 200 on
# https://dashscope-intl.aliyuncs.com/compatible-mode/v1/chat/completions
# with reasoning_content populated (thinking mode default on). The
# mainland endpoint `dashscope.aliyuncs.com` returns 401 invalid_api_key
# — the provisioned key is intl-bound. Failover policy documented in
# ops/litellm/README.md §Sprint-10-Task-1.4.
# refs sprint-10-brief §1.4
- model_name: qwen-max
litellm_params:
model: openrouter/qwen/qwen-max
api_key: os.environ/OPENROUTER_API_KEY
# ── Qwen3.6-35B-A3B canonical + dual-route (primary + fallback) ──
#
# Canonical alias routes to DashScope-intl direct as primary. The
# 3.6-35B slug IS live on the intl tenant (verified 2026-04-21 Sprint
# 10 Task 1.4 smoke). Earlier comment asserting the slug was "not live
# on any public API" was wrong on the intl tenant — only the mainland
# DashScope and OpenRouter catalogs lacked the exact slug.
#
# Callers should prefer the canonical `qwen3.6-35b-a3b` alias — it
# routes to DashScope-intl native, giving us the real 3.6 model
# rather than the Qwen3.5-35B-A3B OpenRouter fallback. If DashScope
# rate-limits or is unavailable, caller-side retry should fall back
# to `qwen3.6-35b-a3b-via-openrouter` (bridge route, one-minor regress
# to 3.5 until OR carries 3.6).
- model_name: qwen3.6-35b-a3b
litellm_params:
model: openai/qwen3.6-35b-a3b
api_base: https://dashscope-intl.aliyuncs.com/compatible-mode/v1
api_key: os.environ/DASHSCOPE_API_KEY
# Explicit DashScope alias (same target, stable naming for operator
# scripts that want to pin to DashScope irrespective of canonical
# alias flips).
- model_name: qwen3.6-35b-a3b-via-dashscope
litellm_params:
model: openai/qwen3.6-35b-a3b
api_base: https://dashscope-intl.aliyuncs.com/compatible-mode/v1
api_key: os.environ/DASHSCOPE_API_KEY
# Sprint 12 Task 2 C3 Stage 2 Mini Retry v3 (2026-04-23): -direct suffix
# alias expected by the v3 brief §0.2 alias roster. Same route as the
# `qwen3.6-35b-a3b-via-dashscope` sibling above, but with the Stage 2
# LOCKED config materialized as explicit LiteLLM params:
# - extra_body.enable_thinking: true (per models.json `stage2Config`)
# - max_tokens: 16000 (Mini Retry v3 §2.1 — reduced
# from v2 64000 to avoid DashScope tail-latency timeouts)
# 2026-05-26 fix: 0/15 fail rate at N=160 vs 86.5% Sonnet baseline was
# caused by thinking NOT actually defaulting on through this route —
# verbose-but-imprecise output failed user_message_checker. Forcing
# thinking on at the LiteLLM layer is the durable fix and matches the
# stage2Config metadata that always claimed this behavior.
- model_name: qwen3.6-35b-a3b-via-dashscope-direct
litellm_params:
model: openai/qwen3.6-35b-a3b
api_base: https://dashscope-intl.aliyuncs.com/compatible-mode/v1
api_key: os.environ/DASHSCOPE_API_KEY
max_tokens: 16000
extra_body:
enable_thinking: true
# OpenRouter bridge — failover when DashScope rate-limits or is
# unavailable. Note: OpenRouter's Qwen catalog does NOT have the exact
# 3.6-35b-a3b slug — this route lands on qwen3.5-35b-a3b, so it's a
# one-minor-version regression. Acceptable for transient failover,
# not for primary use.
- model_name: qwen3.6-35b-a3b-via-openrouter
litellm_params:
model: openrouter/qwen/qwen3.5-35b-a3b
api_key: os.environ/OPENROUTER_API_KEY
# Sibling alias that points to Qwen3.5-35B-A3B directly (no ambiguity).
- model_name: qwen3.5-35b-a3b
litellm_params:
model: openrouter/qwen/qwen3.5-35b-a3b
api_key: os.environ/OPENROUTER_API_KEY
# ── Mistral ────────────────────────────────────────────────────
- model_name: mistral-large-latest
litellm_params:
model: mistral/mistral-large-latest
api_key: os.environ/MISTRAL_API_KEY
- model_name: mistral-small-latest
litellm_params:
model: mistral/mistral-small-latest
api_key: os.environ/MISTRAL_API_KEY
- model_name: codestral-latest
litellm_params:
model: mistral/codestral-latest
api_key: os.environ/MISTRAL_API_KEY
# ── MiniMax ────────────────────────────────────────────────────
# `abab6.5s-chat` was deprecated by MiniMax (confirmed 2026-04-21 via
# 400 "unknown model" response). Repointed to MiniMax-M1, the nearest
# current successor. The `minimax-01` alias is preserved for any
# callers using the provider.ts UI slug.
- model_name: minimax-01
litellm_params:
model: openai/MiniMax-M1
api_base: https://api.minimaxi.chat/v1
api_key: os.environ/MINIMAX_API_KEY
- model_name: minimax-2.7
litellm_params:
model: openai/MiniMax-Text-01
api_base: https://api.minimaxi.chat/v1
api_key: os.environ/MINIMAX_API_KEY
# ── GLM / Zhipu ────────────────────────────────────────────────
- model_name: glm-4-plus
litellm_params:
model: openai/glm-4-plus
api_base: https://open.bigmodel.cn/api/paas/v4
api_key: os.environ/ZHIPU_API_KEY
- model_name: glm-5
litellm_params:
model: openai/glm-5
api_base: https://open.bigmodel.cn/api/paas/v4
api_key: os.environ/ZHIPU_API_KEY
# ── Tavily (search — not LLM, but here for completeness) ───────
# Tavily is consumed by Waggle search routes, not LiteLLM.
# ── Added for PromptAssembler eval (v4) ────────────────────────
# Spec: docs/specs/PROMPT-ASSEMBLER-V4.md §11.2
# Claude Opus 4.7 — native Anthropic. Plain family alias verified via
# /v1/models probe 2026-04-17 (brief's speculative date suffix 20260416
# returned 404; plain alias resolves).
- model_name: claude-opus-4-7
litellm_params:
model: anthropic/claude-opus-4-7
api_key: os.environ/ANTHROPIC_API_KEY
# Gemma 4 31B Dense — OpenRouter (verified live 2026-04-17)
- model_name: gemma-4-31b
litellm_params:
model: openrouter/google/gemma-4-31b-it
api_key: os.environ/OPENROUTER_API_KEY
# Gemma 4 26B MoE (3.8B active) — OpenRouter
# NOTE: live OpenRouter slug is `google/gemma-4-26b-a4b-it` (a4b middle)
# vs brief's suggested `gemma-4-26b-it`. Substitution logged 2026-04-17.
- model_name: gemma-4-26b-moe
litellm_params:
model: openrouter/google/gemma-4-26b-a4b-it
api_key: os.environ/OPENROUTER_API_KEY
# Qwen3-30B-A3B (30B total / 3B active) — OpenRouter
# NOTE: live slug is `qwen/qwen3-30b-a3b-instruct-2507` (date-tagged)
# vs brief's suggested `qwen3-30b-a3b-instruct`. Substitution logged.
- model_name: qwen3-30b-a3b
litellm_params:
model: openrouter/qwen/qwen3-30b-a3b-instruct-2507
api_key: os.environ/OPENROUTER_API_KEY
# ── Qwen3.6-35B-A3B LOCKED decision (2026-04-19) ──────────────
# Canonical engine for Waggle Pro/Teams default + KVARK prod +
# Track 2 benchmarks (H-42 / H-43 / H-44). Per the owner's LOCKED decision
# of 2026-04-19.
# 35B total / 3B active MoE, Apache-2.0, 262K native / 1M YaRN.
# Thinking mode default ON, togglable via `enable_thinking: false`.
# HF source (verified 2026-04-19):
# https://huggingface.co/Qwen/Qwen3.6-35B-A3B
#
# The canonical `qwen3.6-35b-a3b` routing ENTRY lives above in the
# Alibaba/Qwen section (next to its dashscope + openrouter siblings).
# DashScope-intl direct is primary as of Sprint 10 Task 1.4
# (2026-04-21). This comment block stays here as the LOCKED decision
# memo anchor — do not duplicate the route entry.
# ── Added for PromptAssembler v5 eval ──────────────────────────
# Spec: v5 brief §11.2. Entries verified live via provider docs on
# 2026-04-17. Eval harness runs a slug probe at start and logs any
# substitutions. Cost is deliberately high for this run — v5 uses a
# 4-model frontier judge ensemble to avoid same-family judge bias
# from v4 (Sonnet 4.6 judging Claude candidates).
# ── Judge panel (4 frontier models, 4 different labs) ──
# Gemini 3.1 Pro — Google's frontier (released Feb 19, 2026)
# Current SOTA on GPQA Diamond (94.3%), LMArena #1 (1501 Elo).
# Neutral to all candidate families.
#
# Manifest v5 §5.2 addendum (2026-04-24, Stage 3 rekick): rpm: 20 applied
# after §1.3 Gate P+ probe FAIL (24/50 HTTP 429) revealed Google per-model
# preview cap of 25 RPM on generate_requests_per_model for this upstream.
# 20 < 25 Google cap with 5 RPM margin for burst variance. This is the
# single permitted §11 code-freeze exception in v5 per PM P4 path. See
# benchmarks/preregistration/manifest-v5-preregistration.md §5.2 and §11.
- model_name: gemini-3.1-pro
litellm_params:
model: gemini/gemini-3.1-pro-preview
api_key: os.environ/GEMINI_API_KEY
rpm: 20
# Sprint 12 Task 2 C3 Stage 2 Mini Retry v3 (2026-04-23): -preview suffix
# alias expected by the v3 brief §0.2 alias roster (judge_tie_breaker).
# Same upstream as `gemini-3.1-pro` above — alias name encodes the
# "-preview" provenance explicitly for manifest audit clarity.
#
# Manifest v5 Fold-in 3.5b (2026-04-24): defensive-mirror rpm: 20 applied
# here as well because this alias shares the same Google upstream bucket
# (`gemini/gemini-3.1-pro-preview` at 25 RPM per-model cap) with the
# primary `gemini-3.1-pro` alias (edited in Step 2). Isolation sweep
# found non-zero (dormant) references to this alias in repo-wide grep,
# so defensive posture per PM brief Fold-in 3.5b decision tree.
- model_name: gemini-3.1-pro-preview
litellm_params:
model: gemini/gemini-3.1-pro-preview
api_key: os.environ/GEMINI_API_KEY
rpm: 20
# GPT-5.4 — OpenAI's flagship (released Mar 5, 2026)
# Strong judgment on structured outputs. Neutral to all candidates.
- model_name: gpt-5.4
litellm_params:
model: openai/gpt-5.4
api_key: os.environ/OPENAI_API_KEY
# GPT-5.4 Pro — stronger variant. Held in reserve as ensemble fallback
# per brief §11.2 ("GPT-5.4 Pro optional fallback"). Not called by default.
- model_name: gpt-5.4-pro
litellm_params:
model: openai/gpt-5.4-pro
api_key: os.environ/OPENAI_API_KEY
# Grok 4.20 — xAI frontier with reasoning toggle (released Mar 31, 2026)
# 2M context, lowest hallucination rate on market. Fully outside
# Claude/Qwen/Gemma training lineage.
- model_name: grok-4.20
litellm_params:
model: xai/grok-4.20
api_key: os.environ/XAI_API_KEY
# MiniMax M2.7 — SOTA agentic reasoning (released Apr 2026)
# 56.22% SWE-Pro, beats Opus 4.6 on agentic benchmarks. Chinese lab,
# maximally different training lineage from Western frontier. Per
# brief: the existing `minimax-2.7` alias above points to the older
# MiniMax-Text-01 endpoint — this new entry routes to the M2.7 API.
# Note: MiniMax uses interleaved thinking; keep <think>...</think>
# tags in message history to preserve scoring quality.
- model_name: minimax-m2.7
litellm_params:
model: openai/MiniMax-M2.7
api_base: https://api.minimaxi.chat/v1
api_key: os.environ/MINIMAX_API_KEY
# ── Manifest v6 judge ensemble swap (2026-04-24) ──────────────────
# Added under v6 Phase 1 Commit 2 authority on PM ratification of
# §1.3g-§1.3h-C sequence closure. v6 supersedes v5 §11 freeze on
# litellm-config.yaml (single permitted amendment; Gemini rpm:20
# entry retained above as orphan audit artefact).
# Anchor: benchmarks/preregistration/manifest-v6-preregistration.md
# v6 selection rationale documented in §0.5 delta log (MD) + judge_ensemble
# block (YAML). All v5 aliases retained; v6 additions are orthogonal.
# v6 primary_judge_3 — MiniMax M2.7 via OpenRouter.
# §1.3h verified: direct routes (api.minimaxi.com + api.minimax.chat v2)
# fail 401 on MINIMAX_GROUP_ID + Bearer auth; OR fallback 7/7 parse.
# OR catalog slug verified live 2026-04-24: minimax/minimax-m2.7
# @ 196K ctx, $0.30 prompt / $1.20 completion per 1M tokens.
- model_name: minimax-m27-via-openrouter
litellm_params:
model: openrouter/minimax/minimax-m2.7
api_key: os.environ/OPENROUTER_API_KEY
# v6 backup_judge — Kimi K2.6 via Moonshot direct.
# §1.3g verified: direct route (api.moonshot.ai/v1, intl tenant) stable.
# Moonshot catalog verified live 2026-04-24: kimi-k2.6.
# Per-instance failover on primary_judge_3 (MiniMax) failure per v6 §5.2
# (API error / parse failure / 60s timeout). If both fail same instance
# -> judge_ensemble_fail (evaluator_loss in denominator, §9).
# Routed as OpenAI-compatible via api_base override (same pattern as
# kimi-k2-native above); LiteLLM's moonshot/ provider prefix not used
# because the openai/ prefix + api_base is proven stable per v5 config.
- model_name: kimi-k26-direct
litellm_params:
model: openai/kimi-k2.6
api_base: https://api.moonshot.ai/v1
api_key: os.environ/MOONSHOT_API_KEY
# ── Small-model additions for v5 eval (candidate, not judge) ──
# Qwen3-30B-A3B Thinking variant — dedicated reasoning mode.
# v4 used the instruct variant above; v5's H condition tests this
# thinking variant to measure whether reasoning mode flips PA's
# benefit direction (the publishable claim if it does).
# OpenRouter slug verified 2026-04-17: 131K context.
- model_name: qwen3-30b-a3b-thinking
litellm_params:
model: openrouter/qwen/qwen3-30b-a3b-thinking-2507
api_key: os.environ/OPENROUTER_API_KEY
# ── Genspark ───────────────────────────────────────────────────
# Added 2026-04-21. Genspark's gsk-* token is a JWT whose payload
# references `claude_big_model` / `claude_middle_model` / `claude_small_model`
# fields — i.e. they run a Claude-compatible proxy. Endpoint slug below
# is a best-effort based on their public docs; validate with a smoke
# call after first use. The user's key is in vault + .env as
# GENSPARK_API_KEY but this route is NOT wired into Waggle UI yet.
- model_name: genspark-claude
litellm_params:
model: anthropic/claude-sonnet-4-5-20250929
api_base: https://api.genspark.ai/anthropic/v1
api_key: os.environ/GENSPARK_API_KEY
litellm_settings:
drop_params: true
request_timeout: 120