govai 2.4

.gitattributes

@@ -49,3 +49,4 @@ desktop/govai-setup-1.0.0.exe filter=lfs diff=lfs merge=lfs -text
 desktop/govai-setup-1.0.0.exe.blockmap filter=lfs diff=lfs merge=lfs -text
 desktop/govai-setup-1.0.1.exe filter=lfs diff=lfs merge=lfs -text
 desktop/govai-setup-1.0.1.exe.blockmap filter=lfs diff=lfs merge=lfs -text
+govai-2.4-q4_k_m.gguf filter=lfs diff=lfs merge=lfs -text

README.md

@@ -14,6 +14,8 @@ tags:
 - chat
 - assistant
 - reasoning
+- tool-calling
+- function-calling
 - on-device
 - offline
 - education
@@ -23,7 +25,7 @@ tags:
 
 <img src="https://huggingface.co/govailabs/govai/resolve/main/assets/govai-mark.png" alt="govai" width="96">
 
-<h1 align="center">govai 2.3</h1>
+<h1 align="center">govai 2.4</h1>
 
 **A small, fast AI assistant that runs on your own computer.** Free, open and private.
 
@@ -34,12 +36,15 @@ tags:
 </div>
 
 govai answers in plain words, writes full step-by-step guides when you ask for detail, solves maths
-and science problems, and can think before answering hard questions. One 1.3 GB file, fully offline.
+and science problems, and can think before answering hard questions. **New in 2.4: tool calling.**
+Give govai tools (search your files, a calculator, any API) and it decides by itself when to call
+one, and answers directly when none fits. One 1.3 GB file, fully offline.
 
 | | |
 |---|---|
-| **Size** | ~2B parameters · `govai-2.3-q4_k_m.gguf` · 1.31 GB (Q4_K_M) |
+| **Size** | ~2B parameters · `govai-2.4-q4_k_m.gguf` · 1.31 GB (Q4_K_M) |
 | **Context** | 16,384 tokens recommended |
+| **Tools** | OpenAI-style function definitions (Ollama `tools`, llama.cpp `--jinja`) |
 | **Languages** | English (best), Hindi, Telugu |
 | **Speed** | ~150 tokens/s on a gaming GPU; usable on CPU |
 | **License** | Apache 2.0 · by govai labs |
@@ -49,20 +54,30 @@ and science problems, and can think before answering hard questions. One 1.3 GB
 All scores: the Q4_K_M file above, temperature 0, unfinished answers counted as wrong, and no test
 question used in training (8-word overlap check).
 
-![Thinking on vs off](https://huggingface.co/govailabs/govai/resolve/main/charts/thinking.png)
+![Tool calling](https://huggingface.co/govailabs/govai/resolve/main/charts/tools.png)
 
-| Same questions both ways | Thinking off | Thinking on |
+| Tool calling | govai 2.3 | govai 2.4 |
 |---|---|---|
-| GSM8K · school maths (250) | **82.8%** | 80.8% |
-| ARC-Challenge · science (300) | 83.0% | **85.0%** |
-| MMLU · knowledge (400) | 61.5% | **69.2%** |
+| BFCL v3 · one call (100) | 21% | **85%** |
+| BFCL v3 · pick the right tool (100) | 38% | **85%** |
+| BFCL v3 · several calls at once (100) | 6% | **78%** |
+| BFCL v3 · no tool fits, don't call (100) | **95%** | 85% |
+| **BFCL v3 · average** | 40.0% | **83.2%** |
+| Your files: right tool (search / list / read, 20) | 12/20 | **20/20** |
+| Your files: answer uses what the tool returned (20) | 8/20 | **19/20** |
 
-Thinking helps most on knowledge (+7.7 points) and science (+2.0); for school maths, direct answers
-are already as good.
+The file test uses made-up computers; half of its questions use file kinds that never appeared in
+training (insurance policy, salary slip, blood test).
 
-![Full benchmark sets](https://huggingface.co/govailabs/govai/resolve/main/charts/benchmarks.png)
+![govai 2.3 to govai 2.4](https://huggingface.co/govailabs/govai/resolve/main/charts/progress.png)
 
-![govai 2 to govai 2.3](https://huggingface.co/govailabs/govai/resolve/main/charts/progress.png)
+| Same questions, thinking off | govai 2.3 | govai 2.4 |
+|---|---|---|
+| GSM8K · school maths (250) | 82.8% | 82.4% |
+| ARC-Challenge · science (300) | 83.0% | 83.3% |
+| MMLU · knowledge (400) | 61.5% | 61.8% |
+
+Maths, science and knowledge stay level with govai 2.3 (differences are within ±5 points).
 
 ![Everyday chat quality](https://huggingface.co/govailabs/govai/resolve/main/charts/chat_quality.png)
 
@@ -73,13 +88,32 @@ ollama run hf.co/govailabs/govai:Q4_K_M          # Ollama
 llama-cli -hf govailabs/govai --jinja            # llama.cpp
 ```
 
+<details>
+<summary>Tool calling (Ollama, Python)</summary>
+
+```python
+import ollama
+
+tools = [{"type": "function", "function": {
+    "name": "search_files",
+    "description": "Search the user's files by words in the file name.",
+    "parameters": {"type": "object", "properties": {"query": {"type": "string"}}, "required": ["query"]}}}]
+
+reply = ollama.chat(model="hf.co/govailabs/govai:Q4_K_M", think=False, tools=tools,
+                    messages=[{"role": "user", "content": "Where is my resume?"}])
+print(reply.message.tool_calls)   # [search_files(query="resume")]
+# Run the tool, add {"role": "tool", "content": <result>} to the messages, and ask again for the answer.
+```
+
+</details>
+
 <details>
 <summary>Python (llama-cpp-python)</summary>
 
 ```python
 from llama_cpp import Llama
 
-llm = Llama.from_pretrained(repo_id="govailabs/govai", filename="govai-2.3-q4_k_m.gguf", n_ctx=16384)
+llm = Llama.from_pretrained(repo_id="govailabs/govai", filename="govai-2.4-q4_k_m.gguf", n_ctx=16384)
 reply = llm.create_chat_completion(
     messages=[
         {"role": "system", "content": "You are govai, a friendly and helpful AI assistant. Answer clearly in simple words."},
@@ -92,9 +126,9 @@ print(reply["choices"][0]["message"]["content"])
 
 </details>
 
-**Settings:** temperature 0.3 (0 for maths), top_p 0.9, top_k 20. Turn thinking on (`think: true` in
-Ollama) for hard problems and give it room (4,096+ tokens). Ask "explain fully" or "step by step"
-for a complete guide; short questions get short answers.
+**Settings:** temperature 0.3 (0 for maths and tool calls), top_p 0.9, top_k 20. Turn thinking on
+(`think: true` in Ollama) for hard problems and give it room (4,096+ tokens). Ask "explain fully" or
+"step by step" for a complete guide; short questions get short answers.
 
 ## govai for Windows
 
@@ -109,6 +143,8 @@ SHA-256: `000b2da90830d793e2ec9b5a2f5e7adf0b0dcde6d0b3c26a17cdddc4b08b4d31`
 
 - A small model: it can get facts wrong, especially about people, places, dates and recent events.
   The govai app searches the web for those; on its own, check important facts.
+- Tool calls can still be wrong or missing an argument; check what a tool will do before running it,
+  and give govai read-only tools where you can.
 - Hindi and Telugu answers are weaker than English ones.
 - Very hard problems (competition maths, graduate science) need thinking on and a long budget.
 - Not for professional medical, legal or financial advice.
@@ -122,7 +158,7 @@ Apache 2.0 (`LICENSE`), provided "as is". Open-source notices are in `NOTICE`.
 
 ```bibtex
 @misc{govai2026,
-  title  = {govai 2.3: a small, fast, private AI chat assistant},
+  title  = {govai 2.4: a small, fast, private AI assistant with tool calling},
   author = {govai labs},
   year   = {2026},
   url    = {https://huggingface.co/govailabs/govai}

charts/benchmarks.png → govai-2.4-q4_k_m.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e1fb8d34477aeac8d8f3020b2771a8495bc9151f26b6104ba551c345001fb944
-size 111684
+oid sha256:c4e87866dc8410d9a0d4424cbdb098b3b5083a243642cfaebe51e752ff759bfe
+size 1312164384

latest.json

@@ -1,8 +1,8 @@
 {
   "name": "govai",
-  "version": "2.3",
-  "file": "govai-2.3-q4_k_m.gguf",
-  "size": 1312164480,
-  "sha256": "dc18c3dc28d340986c749f5d6f008e65379825de344cd660591f20eb4795712c",
-  "url": "https://huggingface.co/govailabs/govai/resolve/main/govai-2.3-q4_k_m.gguf"
+  "version": "2.4",
+  "file": "govai-2.4-q4_k_m.gguf",
+  "size": 1312164384,
+  "sha256": "c4e87866dc8410d9a0d4424cbdb098b3b5083a243642cfaebe51e752ff759bfe",
+  "url": "https://huggingface.co/govailabs/govai/resolve/main/govai-2.4-q4_k_m.gguf"
 }