Otter Models

Otter Models

chimera otter accepts any model id understood by Chimera’s provider factory (chimera/providers/factory.py). The provider chain auto-detects the right backend from the prefix (claude-* → Anthropic, gpt-* / o1* / o3* → OpenAI, gemini-* → Google, glm-* / kimi-* / qwen* / llama* / mistral* / phi* → Ollama, vendor/name → OpenRouter when the OpenRouter key is set). Otter does not maintain a parallel registry; what works for chimera mink works here.

This page is the otter-flavored entry point for picking a model:

• The default the resolver lands on when you don’t pass --model.
• The OTTER_MODEL environment variable for pinning a model in CI / a user’s shell.
• Concrete ids that are tested today.

For the smoke-tested matrix of Ollama tags with measured cold-start, tool turn time, and per-turn cost, see docs/mink/models.md; those numbers transfer to otter unchanged.

Default model

Otter’s resolver picks a default in this order (chimera/otter/providers.py:_resolve_model):

1. --model <id> on the CLI (highest precedence).
2. $OTTER_MODEL environment variable.
3. If $ANTHROPIC_API_KEY is set: claude-sonnet-4-6.
4. If $OPENROUTER_API_KEY is set: anthropic/claude-sonnet-4.
5. If $OPENAI_API_KEY is set: gpt-4o.
6. Otherwise: a friendly error pointing at the three env vars above.

So a fresh laptop with only $ANTHROPIC_API_KEY exported gets claude-sonnet-4-6 automatically. A CI runner with $OPENROUTER_API_KEY falls back to anthropic/claude-sonnet-4 (routed through the OpenAI-compatible adapter). A purely-OpenAI environment lands on gpt-4o.

The OTTER_MODEL env var

OTTER_MODEL is the project’s escape hatch for “I want this model every time, but I don’t want to pass --model.” It sits between the explicit --model flag and the provider-specific defaults:

export OTTER_MODEL=claude-sonnet-4-6
chimera otter -p "summarize this repo"                  # uses claude-sonnet-4-6
chimera otter --model gpt-4o -p "draft a release note"  # overrides to gpt-4o

A few practical patterns:

# CI pins the cheapest first-class model.
export OTTER_MODEL=anthropic/claude-haiku-4
export OPENROUTER_API_KEY=sk-or-...

# Local dev pins a long-context model for refactor sessions.
export OTTER_MODEL=kimi-k2.6:cloud

# Privacy-sensitive shell pins a local Ollama tag.
export OTTER_MODEL=qwen3:32b

Compared to mink: mink has both $CHIMERA_MINK_MODEL and $CHIMERA_MINK_FALLBACK (the second is used when the primary tag is missing locally). Otter ships only $OTTER_MODEL — the fallback story is handled by the resolver chain (Anthropic → OpenRouter → OpenAI), and unknown Ollama tags surface as a provider error rather than silently swapping.

Recommended models

The list below names ids that have been smoke-tested against chimera otter on at least one developer machine. Absence from this list does not mean “unsupported”; it just means we haven’t measured it.

Anthropic

Model id	Notes
claude-sonnet-4-6	Otter default. Strongest tool calling in the Anthropic line; supports extended thinking, prompt caching, vision.
claude-opus-4	Strongest reasoning; 2x cost of sonnet. Good for long-form refactors.
claude-haiku-4	Cheapest first-class Anthropic model; fine for short turns and quick edits.
claude-sonnet-3-7	Older release; useful when reproducing historical benchmark runs.

OpenAI

Model id	Notes
gpt-4o	OpenAI default in the otter resolver. Streaming, tool calls, vision, JSON mode.
gpt-4o-mini	Cheaper; suitable for one-shots and CI smoke tests.
o3	Reasoning-tuned; spends more on reasoning tokens. The provider tracks them via chimera/providers/cost_tracker.py.
o3-mini	Smaller reasoning model; fast for proofs / structured tasks.

OpenRouter (vendor/name)

OpenRouter is keyed off the / in the model id. Anything with a slash, when $OPENROUTER_API_KEY is set, routes through the OpenAI-compatible adapter at https://openrouter.ai/api/v1.

Model id	Notes
anthropic/claude-sonnet-4	OpenRouter default in the otter resolver.
anthropic/claude-opus-4	Same family, stronger reasoning.
google/gemini-2.5-pro	Long context, strong code-gen.
meta-llama/llama-3.3-70b	Open-weight; cheapest of the three.
qwen/qwen-2.5-coder-32b	Open-weight code model; fast on OpenRouter.
deepseek/deepseek-chat	Inexpensive general model with tool calls.

Ollama (local + cloud tags)

Model id	Kind	Context	Notes
glm-5.1:cloud	cloud	131072	Validated working baseline; native tool calls.
glm-5:cloud	cloud	131072	Fast text response, dispatches tools cleanly.
kimi-k2.6:cloud	cloud	262144	Reasoning model (think:true auto-enabled). Long context.
kimi-k2.5:cloud	cloud	131072	Predecessor; steady tool use.
minimax-m2.7:cloud	cloud	131072	Newer minimax weights.
qwen3.5:cloud	cloud	131072	Fast, clean tool dispatch.
gpt-oss:120b-cloud	cloud	131072	Open-weight 120B; fast tool turn.
qwen3:32b	local	131072	Local fallback; runs on a 24 GB GPU.
llama3.1:70b-instruct	local	131072	Heavier local; good on 2× GPU rigs.

For per-tag cold-start, tool dispatch wall-clock, and per-turn cost (all measured against a real Ollama daemon) see docs/mink/models.md.

Google

Model id	Notes
gemini-2.5-pro	1M context, strong code-gen; non-streaming via the chimera adapter.
gemini-2.0-flash	Fastest Gemini; good for cheap one-shots.

Anthropic-compatible third-party endpoints

Set ANTHROPIC_BASE_URL + ANTHROPIC_AUTH_TOKEN and the Anthropic provider talks to whatever endpoint you point it at:

Endpoint	Model ids
api.z.ai/v1/anthropic (GLM)	glm-4.6, glm-4.5
Moonshot Anthropic-compat	kimi-k2-instruct
In-house gateway	whatever your gateway exposes

Selecting a model

Three layered ways to pick a model — last write wins:

1. CLI flag (highest): chimera otter --model <id>.
2. Environment: export OTTER_MODEL=<id> in the shell or CI.
3. Default chain: the first env var that’s set among $ANTHROPIC_API_KEY / $OPENROUTER_API_KEY / $OPENAI_API_KEY.

Examples:

# One-off override.
chimera otter --model claude-opus-4 -p "review this PR"

# Pin for the shell session.
export OTTER_MODEL=anthropic/claude-sonnet-4
export OPENROUTER_API_KEY=sk-or-...
chimera otter -p "summarize"

# Resolver default (no flag, no env, but $ANTHROPIC_API_KEY set).
export ANTHROPIC_API_KEY=sk-ant-...
chimera otter -p "explain this repo"   # → claude-sonnet-4-6

REPL /model slash command

The otter REPL exposes a /model slash command. With a single-model launch it prints the active model. With --models a,b,c it cycles through the comma-separated list:

chimera otter --models claude-sonnet-4-6,gpt-4o,gemini-2.5-pro
# inside the REPL:
/model            # show current
/model next       # cycle to gpt-4o
/model prev       # cycle back

Each cycle rebuilds the provider behind the live session; sessions event-log records the model on each turn so sessions show can show you which model produced which event.

Adding your own model

Three paths, depending on how the model is served:

1. Custom Ollama tag. Build a Modelfile in front of a base tag, ollama create my-agent -f ./Modelfile, then chimera otter --model my-agent -p "...". The provider hits /api/chat like any other Ollama tag.
2. Custom OpenAI-compatible endpoint. Set OPENAI_BASE_URL (or pass base_url= when you build the provider directly) and use a model id the endpoint understands. Otter routes through the compatible adapter when the prefix doesn’t match a built-in.
3. Brand-new provider. Implement Provider (chimera/providers/base.py) and call register_provider (chimera/providers/registry.py). See the providers.md “Custom providers” section for the factory contract.

Cost accounting

Otter writes cost_usd to each run’s summary.json (chimera/otter/cli.py:_write_run_summary). The number comes from chimera/providers/cost.py (per-model pricing) and the cost tracker (chimera/providers/cost_tracker.py). Models without an entry in the pricing table report 0.0 — that’s not free, it’s “we don’t know what the upstream charges.” Hosted models served by the local Ollama daemon also land at 0.0 because /api/chat does not surface a price field.

To audit your spend across persisted runs:

chimera otter sessions list --json | jq '[.[].cost_usd] | add'

A roll-up subcommand similar to chimera mink runs cost is on the roadmap.

See also

• providers.md — provider chain that picks the SDK to drive a given model id.
• docs/mink/models.md — measured smoke matrix for every Ollama tag, with run ids and cost data.
• quickstart.md — first-call walkthrough.
• sessions.md — where the model id lands in summary.json.