Skip to content
Chimera

Mink Model Compatibility

Mink Model Compatibility

Section titled “Mink Model Compatibility”

This page lists every Ollama model that has been end-to-end smoke-tested against chimera mink on this machine, along with the captured run IDs that produced the numbers. Every row is reproducible from the eventlog directories cited.

The two smokes used are identical for every model:

SmokePurposeCommand (paraphrased)
A: text-onlyproves provider wiringmink -p "Say hello in three words." --max-steps 1
B: tool dispatchproves the model emits a real tool_calls block that mink can route to the bash toolmink -p "Use the bash tool to run 'echo HELLO_FROM_<MODEL>' and report the result." --max-steps 4

Both invoked with --permission-mode bypassPermissions --output-format json. “Tool calls” below refers to tool_calls_total from the per-run ~/.chimera/eventlog/<run_id>/summary.json. Wall-clock is ended_at - started_at from the same file. Costs are whatever OllamaProvider reported back to the run summary; cloud models that the local Ollama daemon does not price out come back as $0.0000 (the cost is borne by the upstream account, not surfaced through /api/chat).

Section titled “Recommended models”
ModelKindContext (default)Tool calls observedSmoke A turnSmoke B turnCost / turn (B)Notes
glm-5.1:cloudcloud131072130.0s17.0s$0.0204Validated working baseline. Slowest cold-start in this batch (30s on first call).
glm-5:cloudcloud13107211.0s9.0s$0.0214Fastest text response, dispatches tools cleanly. Cheap-ish per turn.
kimi-k2.5:cloudcloud13107214.0s17.0s$0.0000Reasoning model (think:true auto-enabled by provider). Steady tool use.
kimi-k2.6:cloudcloud13107219.0s11.0s$0.0000Default mink model. Reasoning enabled by default.
minimax-m2:cloudcloud13107211.0s9.0s$0.0000Tied for fastest tool turn.
minimax-m2.7:cloudcloud13107212.0s10.0s$0.0000Successor of m2; same tool-use behavior.
qwen3.5:cloudcloud13107213.0s10.0s$0.0000Clean tool dispatch, fast.
gpt-oss:120b-cloudcloud13107211.0s9.0s$0.0000Tied for fastest tool turn. One transient 500 from /api/chat on first try; succeeded on retry.
llama3.2:3blocal131072122.0s (text smoke failed)*49.0s$0.0000Smallest viable local model. Slow and unreliable on free-form text smoke; surprisingly does dispatch the bash tool.

*Smoke A returned success=false (“Max steps reached”) for llama3.2:3b. See “Models that failed or are partial” below.

Per-model details

Section titled “Per-model details”

Setup commands assume Ollama is running at http://localhost:11434. Cloud tagged models (*:cloud) are manifest-only pulls (a few hundred bytes each) and require ollama signin to be authenticated against an account that has access to the cloud model catalog. Local models are full weight downloads.

glm-5.1:cloud

Section titled “glm-5.1:cloud”
  • Setup: ollama signin (once), then ollama pull glm-5.1:cloud.
  • Strengths: Stable, accurate tool dispatch, tracks the prompt closely.
  • Weaknesses: Saw a 30s cold-start on the first text request of the session; subsequent calls are sub-10s.
  • Suggested flags: --max-steps 50 (default), --tool-timeout 120.
  • Evidence:
    • Smoke A: ~/.chimera/eventlog/mink-20260425T182550-523915ec/summary.json
    • Smoke B: ~/.chimera/eventlog/mink-20260425T183450-71742dd6/summary.json

glm-5:cloud

Section titled “glm-5:cloud”
  • Setup: ollama pull glm-5:cloud.
  • Strengths: Fastest text response in this batch (1.0s), good tool formatting.
  • Weaknesses: Slightly higher per-turn cost surfaced through the provider ($0.0214 on smoke B).
  • Suggested flags: --max-steps 50, --tool-timeout 120.
  • Evidence:
    • Smoke A: ~/.chimera/eventlog/mink-20260425T182630-1a0e2683/summary.json
    • Smoke B: ~/.chimera/eventlog/mink-20260425T183512-430cc00a/summary.json

kimi-k2.5:cloud

Section titled “kimi-k2.5:cloud”
  • Setup: ollama pull kimi-k2.5:cloud.
  • Strengths: Reasoning model; the Ollama provider auto-sets think:true for any tag starting with kimi. Good for multi-step planning.
  • Weaknesses: ~17s on smoke B is mid-pack; reasoning tokens add latency.
  • Suggested flags: --max-steps 50, --tool-timeout 180 (give reasoning room).
  • Evidence:
    • Smoke A: ~/.chimera/eventlog/mink-20260425T182631-dce51214/summary.json
    • Smoke B: ~/.chimera/eventlog/mink-20260425T183521-d621617c/summary.json

kimi-k2.6:cloud

Section titled “kimi-k2.6:cloud”
  • Setup: ollama pull kimi-k2.6:cloud. This is the mink default (see _DEFAULT_MODEL in chimera/mink/cli.py).
  • Strengths: Reasoning enabled by default; well-tuned for the Chimera loop; this is the model the rest of the project benchmarks against.
  • Weaknesses: Cold-start can exceed 2 minutes on a freshly-signed-in account (not observed in this run because the account was already warm).
  • Suggested flags: defaults are tuned for it. --tool-timeout 300 if you expect a cold cloud node.
  • Evidence:
    • Smoke A: ~/.chimera/eventlog/mink-20260425T182651-acb8b23a/summary.json
    • Smoke B: ~/.chimera/eventlog/mink-20260425T183616-d60db3fd/summary.json

minimax-m2:cloud

Section titled “minimax-m2:cloud”
  • Setup: ollama pull minimax-m2:cloud.
  • Strengths: Tied for fastest tool turn (9.0s on smoke B).
  • Weaknesses: None observed in smoke; weights are older than m2.7.
  • Suggested flags: --max-steps 50, --tool-timeout 120.
  • Evidence:
    • Smoke A: ~/.chimera/eventlog/mink-20260425T182704-975eb417/summary.json
    • Smoke B: ~/.chimera/eventlog/mink-20260425T183645-fec8628f/summary.json

minimax-m2.7:cloud

Section titled “minimax-m2.7:cloud”
  • Setup: ollama pull minimax-m2.7:cloud.
  • Strengths: Newer minimax weights; tool dispatch is clean, runs fast.
  • Weaknesses: None observed in smoke.
  • Suggested flags: --max-steps 50, --tool-timeout 120.
  • Evidence:
    • Smoke A: ~/.chimera/eventlog/mink-20260425T182641-5e4c98b3/summary.json
    • Smoke B: ~/.chimera/eventlog/mink-20260425T183544-9f4db463/summary.json

qwen3.5:cloud

Section titled “qwen3.5:cloud”
  • Setup: ollama pull qwen3.5:cloud.
  • Strengths: Fast, clean tool dispatch, very chatty in a good way.
  • Weaknesses: None observed in smoke.
  • Suggested flags: --max-steps 50, --tool-timeout 120.
  • Evidence:
    • Smoke A: ~/.chimera/eventlog/mink-20260425T182643-5be3897c/summary.json
    • Smoke B: ~/.chimera/eventlog/mink-20260425T183554-bb818c61/summary.json

gpt-oss:120b-cloud

Section titled “gpt-oss:120b-cloud”
  • Setup: ollama pull gpt-oss:120b-cloud.
  • Strengths: Tied for fastest tool turn (9.0s). Useful as a third-party baseline for tool-use comparison.
  • Weaknesses: Returned a 500 from /api/chat on the first smoke-B attempt (cold cloud node, transient); succeeded on the immediate retry. Plan for a one-shot retry around this model.
  • Suggested flags: --max-steps 50, --tool-timeout 180. Wrap calls in retry-on-500 if you script against it.
  • Evidence:
    • Smoke A: ~/.chimera/eventlog/mink-20260425T182650-c027e2e2/summary.json
    • Smoke B (retry): ~/.chimera/eventlog/mink-20260425T183635-7391c9e5/summary.json

llama3.2:3b (local)

Section titled “llama3.2:3b (local)”
  • Setup: ollama pull llama3.2:3b.
  • Strengths: Pure local, zero cost, surprisingly does dispatch the bash tool (smoke B succeeded).
  • Weaknesses: Free-form smoke A failed with success=false and "Max steps reached" after 22s — at --max-steps 1 the 3B model spent the budget without producing a stop-able message. Smoke B took 49s, the slowest in the entire batch.
  • Suggested flags: Use as a fallback only. --max-steps 6, --tool-timeout 120. Do not rely on it for production tool loops.
  • Evidence:
    • Smoke A (failed): ~/.chimera/eventlog/mink-20260425T183421-c1be4b5d/summary.json
    • Smoke B (passed): ~/.chimera/eventlog/mink-20260425T183920-8e864999/summary.json

Models that failed or are partial

Section titled “Models that failed or are partial”
ModelStatusReason
glm-4.7-flash (local, 19 GB)failedFirst-load httpx.ReadTimeout even at a 480s overall timeout. The 19GB MoE weights take longer than the default httpx read window to spin up on this hardware. Would need a custom Modelfile that pre-warms with keep_alive raised, or a longer client read timeout in OllamaProvider.
qwen3.5 (local, non-cloud tag)failed to pullError: pull model manifest: 412 ... requires a newer version of Ollama. Upgrade Ollama via https://ollama.com/download to use this tag. The cloud variant qwen3.5:cloud works fine.
llama3.2:3b (local)partialSmoke A failed (Max steps reached); smoke B passed but at 49s. Treat as fallback only.

How to add your own model

Section titled “How to add your own model”

The Ollama provider lives at chimera/providers/ollama.py. Notable defaults:

  • _DEFAULT_NUM_CTX = 131_072 (131k context window pushed to Ollama as num_ctx).
  • _DEFAULT_KEEP_ALIVE = "60m" (model stays warm for an hour after the last call).
  • think=True is auto-enabled when the model tag starts with kimi. Override per-call by passing think=... through provider kwargs.

To register a custom Ollama model:

  1. Create a Modelfile pointing at your base weights and any system-prompt tuning. Example: FROM kimi-k2.6:cloud plus your own SYSTEM.
  2. ollama create my-agent -f ./Modelfile.
  3. chimera mink --model my-agent .... The provider will dispatch through the same /api/chat path, so tool-calling parity is preserved.

For a fully custom provider (e.g., raise num_ctx, change keep_alive, disable thinking on a kimi tag), instantiate OllamaProvider directly in your own script:

from chimera.providers.ollama import OllamaProvider


provider = OllamaProvider(
    model="kimi-k2.6:cloud",
    context_length=200_000,
    keep_alive="2h",
    think=False,
)

Switching models

Section titled “Switching models”

There are two supported ways to switch the model mink talks to:

  1. Per-invocation via --model:

    Terminal window
    chimera mink --model glm-5:cloud -p "..."
    chimera mink --model minimax-m2.7:cloud -p "..."

  2. Per-environment fallback via CHIMERA_MINK_FALLBACK: chimera mink first probes /api/tags for the requested --model. If the tag is not present locally and CHIMERA_MINK_FALLBACK is set, it falls back to that tag instead of failing. Default fallback is qwen3:32b (see chimera/mink/cli.py).

    Terminal window
    export CHIMERA_MINK_FALLBACK=glm-5.1:cloud
    chimera mink --model some-not-yet-pulled-model -p "..."
    # -> uses glm-5.1:cloud

    The legacy CHIMERA_CC_FALLBACK is still accepted with a deprecation warning.

There is no CHIMERA_MINK_MODEL env var; the default model is hardcoded in _DEFAULT_MODEL in chimera/mink/cli.py (currently kimi-k2.6:cloud). If you want a session-wide default, alias it:

Terminal window
alias mink="chimera mink --model glm-5.1:cloud"

Catalog refresh — wave 13 (2026-05)

Section titled “Catalog refresh — wave 13 (2026-05)”

The default ProviderCatalog (chimera/providers/catalog.py) ships explicit ModelConfig bindings for the seven model families below. Routing inference (chimera/providers/factory.py:_infer_provider) is covered by tests/providers/test_catalog_refresh.py.

Model idProviderEndpoint / envPricing (in/out per Mtok)Notes
qwen3-coderollama$OLLAMA_HOST$0 / $0Local Ollama tag (Alibaba). DashScope API path: pass provider_type="compatible" + base_url=https://dashscope-intl.aliyuncs.com/compatible-mode/v1.
qwen3-coder-30bollama$OLLAMA_HOST$0 / $0Same family, 30B coder.
qwen3-32bollama$OLLAMA_HOST$0 / $0Same family, 32B general.
glm-4.6anthropichttps://api.z.ai/api/anthropic + $ANTHROPIC_AUTH_TOKEN$0.6 / $2.2 (placeholder)Zhipu Anthropic-compat. TODO: confirm rates against docs.z.ai.
glm-5.1anthropichttps://api.z.ai/api/anthropic + $ANTHROPIC_AUTH_TOKEN$2 / $8 (mirrors glm-5)Same endpoint as glm-5; pricing TODO until Zhipu publishes 5.1 sheet.
deepseek-v3.1-terminuscompatiblehttps://api.deepseek.com/v1 + $DEEPSEEK_API_KEY$0.27 / $1.10 (placeholder)DeepSeek hosted OpenAI-compat.
deepseek-coder-v3compatiblehttps://api.deepseek.com/v1 + $DEEPSEEK_API_KEY$0.27 / $1.10 (placeholder)Coder line; longest-prefix matched ahead of deepseek-chat.
gpt-oss-120bollama$OLLAMA_HOST$0 / $0OpenAI open weights via Ollama. Routed to ollama by an explicit gpt-oss prefix that fires before the gpt-* → OpenAI rule.
gpt-oss-20bollama$OLLAMA_HOST$0 / $0Smaller OSS sibling.
kimi-k2-0905-previewanthropichttps://api.moonshot.ai/anthropic + $MOONSHOT_API_KEY$0.6 / $2.5 (placeholder)Moonshot Anthropic-compat. :cloud Kimi tags stay served by Ollama.
kimi-k2.5anthropichttps://api.moonshot.ai/anthropic + $MOONSHOT_API_KEY$0.6 / $2.5 (placeholder)Same endpoint, k2.5 GA line.
mistral-codestral-2511ollama$OLLAMA_HOST$0 / $0Mistral coder. For Mistral hosted API, override with provider_type="compatible" + base_url=https://api.mistral.ai/v1 + $MISTRAL_API_KEY.
gemma3-27b-instructollama$OLLAMA_HOST$0 / $0Google open weights. Routed via a new gemma prefix; hosted Gemini stays on the gemini-* → Google branch.

Pricing entries flagged (placeholder) are educated guesses — refresh once the upstream vendor publishes per-SKU rates. Local Ollama tags report $0 because /api/chat does not surface a price field; the real cost is hardware + electricity.

Reproducing this report

Section titled “Reproducing this report”

Every smoke in this report was run against a live Ollama daemon at http://localhost:11434 from this checkout. To reproduce a single row:

Terminal window
# Smoke A
timeout 90 uv run python -m chimera.cli.main mink \
  --model <MODEL> --permission-mode bypassPermissions --output-format json \
  -p "Say hello in three words." --max-steps 1


# Smoke B
timeout 180 uv run python -m chimera.cli.main mink \
  --model <MODEL> --permission-mode bypassPermissions --output-format json \
  -p "Use the bash tool to run 'echo HELLO_FROM_<TAG>' and report the result." \
  --max-steps 4

Each invocation prints a [mink] run saved as mink-<timestamp>-<id> at ~/.chimera/eventlog/mink-<timestamp>-<id>/ line; the summary.json in that directory carries tool_calls_total, cost_usd, started_at, ended_at, and success — the same fields cited above.