An 8B tool-calling model that picks the right tool with valid arguments and does not loop. Fine-tuned from Ministral 3 8B on production agent traffic. Apache-2.0.

Applications

Claude Code ollama launch claude --model OculusMindAI/toolcall-8b-v2

OpenCode ollama launch opencode --model OculusMindAI/toolcall-8b-v2

Hermes Agent ollama launch hermes --model OculusMindAI/toolcall-8b-v2

OpenClaw ollama launch openclaw --model OculusMindAI/toolcall-8b-v2

OculusMind-ToolCall-8B-v2 (“Pico V2”)

Pico V2 is an 8-billion-parameter tool-calling model that runs on a laptop and picks the right tool, with valid arguments, on 95 % of the held-out production turns it was built for — 9 points above its base. It is Mistral’s Ministral-3-8B-Instruct-2512, fine-tuned on production tool-calling data from the agents at OculusMind.AI, and selected from a ten-seed sweep under a hard rule: a seed that ever looped was out, whatever it scored.

ollama run oculusmindai/toolcall-8b-v2

Needs Ollama 0.35.1 or newer (tested on 0.35.1). The download is 9.0 GB, the Q8_0 build. Text only: image input is not supported.

What it does well

Chooses the right tool, with valid arguments, first time. It agrees with production’s tool choice on 95.45 % of held-out tool-turns, against 86.36 % for the base model on the same items.
Does not loop. Degenerate repeated tool calls measured 0.00 % on every population, including a slice built to provoke them, where the base loops on 8.33 %.
Keeps its general ability. It is ahead of the base on GSM8K (87.6 vs 79.7), IFEval (72.2 vs 68.7) and BFCL V4 (69.3 vs 65.2 over all 4,441 items).
Reads long context. It recalls a planted fact and makes a correct tool call at every step from 2K to 250K tokens.

Tool calling

Pass tools in Ollama’s tools array on /api/chat:

curl -s http://127.0.0.1:11434/api/chat -d '{
  "model": "oculusmindai/toolcall-8b-v2",
  "stream": false,
  "messages": [{"role": "user", "content": "What is the weather in San Francisco in celsius?"}],
  "tools": [{
    "type": "function",
    "function": {
      "name": "get_current_weather",
      "description": "Get the current weather for a city",
      "parameters": {
        "type": "object",
        "properties": {
          "city": {"type": "string", "description": "The city name"},
          "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}
        },
        "required": ["city", "unit"]
      }
    }
  }]
}'

It returns a structured call with empty content:

{"role": "assistant", "content": "",
 "tool_calls": [{"function": {"name": "get_current_weather",
   "arguments": {"city": "San Francisco", "unit": "celsius"}}}]}

Send the tool’s result back as a tool message, keeping the same tools array, and it answers: “The current weather in San Francisco is 14°C and foggy.”

Settings in this package

Setting	Value	Why
`temperature`	0	Every published score was measured at 0.
`num_ctx`	32768	Bounds memory. The scores were measured at 65,536 and the window was smoke-tested to 250K; raise it if your machine allows.
`stop`	`</s>`, `[INST]`, `[/INST]`, `[TOOL_RESULTS]`	Without these, a tool call can run on into the next turn and fail to parse.
system prompt	identity only	Replace it with your own agent prompt. The model was fine-tuned on conversations that carried real system prompts and tool schemas.

How the numbers were measured

The scores above come from the Q8_0 GGUF served through llama.cpp, the path the full model card documents. Through Ollama, on the same held-out set, the model scores 93.94 against 95.45 through llama.cpp, with 0 loops on both. The method, every benchmark, the limitations and the file checksums are in the model card: huggingface.co/OculusMindAI/OculusMind-ToolCall-8B-v2.

Licence

Apache-2.0. This package carries the licence and the NOTICE, which states the attribution and modifications that anyone redistributing the model must pass on. Read both with ollama show oculusmindai/toolcall-8b-v2 --license. Fine-tuned from mistralai/Ministral-3-8B-Instruct-2512-BF16 (Apache-2.0).