bobcat is an inference engine optimized for Apple silicon. Run local models on your Mac at unbelievably fast speeds.

bobcat chat -m lfm2.5:2.6b
> What is the capital of Japan? One sentence.
The capital of Japan is Tokyo.

Quick start

Install bobcat on a Mac with Apple silicon and macOS 26 or newer:

curl -fsSL https://raw.githubusercontent.com/jadidbourbaki/bobcat/main/install.sh | sh

Download a model:

bobcat pull lfm2.5:2.6b

Chat with it:

bobcat chat -m lfm2.5:2.6b

Or answer one prompt, for scripts and pipes:

bobcat respond -m lfm2.5:2.6b "Name three primes."

Performance results

LFM2.5-2.6B on an M4 Pro, with a 512-token prompt and 128 generated tokens. The details are here.

bobcat ❤️ agents

Start the server, then point your agent at it:

bobcat serve -m lfm2.5:2.6b -c 65536

Claude Code

ANTHROPIC_BASE_URL=http://127.0.0.1:8080 \
ANTHROPIC_AUTH_TOKEN=bobcat \
ANTHROPIC_API_KEY="" \
CLAUDE_CODE_MAX_CONTEXT_TOKENS=65536 \
claude --model lfm2.5:2.6b

opencode, in opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "bobcat": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "bobcat",
      "options": { "baseURL": "http://127.0.0.1:8080/v1" },
      "models": { "lfm2.5:2.6b": { "name": "LFM2.5-2.6B" } }
    }
  }
}

aider

OPENAI_API_BASE=http://127.0.0.1:8080/v1 \
OPENAI_API_KEY=bobcat \
aider --model openai/lfm2.5:2.6b

bobcat ❤️ decision models

Start the server, then ask it typed questions about any input:

bobcat serve -m qwen3.5:4b

SystemOne

curl http://127.0.0.1:8080/v1/systemone -d '{
  "model": "qwen3.5:4b",
  "state": "Checkout returns 500 errors and orders are blocked.",
  "questions": {
    "team": {"type": "choice", "criteria": {"billing": "Payments", "technical": "Bugs"}},
    "urgency": {"type": "score", "criteria": ["Can wait", "This week", "Today"]},
    "outage": {"type": "noul", "instructions": "Is a service down?"}
  }
}'

TypeSafe SDK

from typesafe_sdk import Choice, TypeSafeClient

client = TypeSafeClient(base_url="http://127.0.0.1:8080", api_key="bobcat")
result = client.system_one(
    "Checkout returns 500 errors.",
    {"team": Choice(criteria={"billing": "Payments", "technical": "Bugs"})},
)

SGLang decisions

curl http://127.0.0.1:8080/v1/decisions -d '{
  "input": "Checkout returns 500 errors.",
  "questions": [{"id": "outage", "type": "yes_no", "question": "A service is down."}]
}'

One decision without a server

bobcat decide -m qwen3.5:4b '{"model": "qwen3.5:4b", "state": "Checkout is down.",
  "questions": {"outage": {"type": "noul", "instructions": "Is a service down?"}}}'

Supported language models

LFM2.5-350M

bobcat pull lfm2.5:350m

LFM2.5-1.2B-Instruct

bobcat pull lfm2.5:1.2b

LFM2.5-2.6B

bobcat pull lfm2.5:2.6b

LFM2.5-8B-A1B

bobcat pull lfm2.5:8b

Qwen3.5-0.8B

bobcat pull qwen3.5:0.8b

Qwen3.5-2B

bobcat pull qwen3.5:2b

Qwen3.5-4B

bobcat pull qwen3.5:4b

Qwen3.5-9B

bobcat pull qwen3.5:9b

Supported decision models

Every language model above answers decisions, except LFM2.5-2.6B, whose chat template always opens a reasoning block before the answer.

Clef-Flash

bobcat pull clef:flash

bobcat is still in alpha. We are rapidly adding support for more models and model families. Please stay tuned!