NVFP4 quantization of BSC-LT/ALIA-40b-fc-2605 packaged for llama.cpp. 27 GB on disk, ~10 tok/s generation on a single NVIDIA GB10 (DGX Spark). This is the function-calling fine-tune of the ALIA-40b base — it emits OpenAI-style <tool_call> JSON blocks when given a tools list at prompt time.

For the vLLM / TensorRT-LLM (compressed-tensors safetensors) version, see montevive/ALIA-40b-fc-2605-NVFP4.

File	Format	Size	Use case
`ALIA-40b-fc-2605.NVFP4.gguf`	NVFP4 (`GGML_TYPE_NVFP4`, type 40)	26.8 GiB	Blackwell GPUs (RTX 50xx, GB10, B-series), llama.cpp

Tool calling

The function-calling capability is the differentiator vs ALIA-40b-instruct-2601. Pass an OpenAI-style tools array in the chat template and the model emits one or more <tool_call>{...}</tool_call> blocks, parseable as JSON with name and arguments keys.

python

from transformers import AutoTokenizer
tok = AutoTokenizer.from_pretrained("BSC-LT/ALIA-40b-fc-2605")

tools = [{
    "type": "function",
    "name": "get_weather",
    "description": "Get current temperature for a given location.",
    "parameters": {
        "type": "object",
        "properties": {"location": {"type": "string", "description": "City and country e.g. 'Madrid, Spain'"}},
        "required": ["location"],
        "additionalProperties": False,
    },
}]

prompt = tok.apply_chat_template(
    [{"role": "user", "content": "What's the weather in Paris?"}],
    tokenize=False, add_generation_prompt=True, tools=tools,
)
# → ...<tool_call>{"name": "get_weather", "arguments": {"location": "Paris, France"}}</tool_call>

In llama.cpp, the --jinja flag picks up the model's embedded chat template (including tools rendering) automatically. The llama-server OpenAI-compatible endpoint exposes tool calling via the standard tools request parameter; clients should parse the <tool_call> blocks from the response or rely on llama.cpp's hermes tool-call parser.

Usage

Requires a recent llama.cpp build. NVFP4 was added in PR #19769 (March 2026); the Llama-architecture Q/K RoPE permute fix needed for correct NVFP4 GGUF output was merged in PR #22611 (May 2026). On Blackwell hardware (compute capability ≥ 12.0), PR #22196 provides native tensor-core acceleration.

bash

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build -DGGML_CUDA=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc)

# Chat with tool calling enabled — --jinja honors the embedded template.
./build/bin/llama-cli \
    -m ALIA-40b-fc-2605.NVFP4.gguf \
    -ngl 99 -c 4096 \
    --jinja -cnv

# Or as an OpenAI-compatible server with WebUI on http://localhost:8080
./build/bin/llama-server \
    -m ALIA-40b-fc-2605.NVFP4.gguf \
    -ngl 99 -c 4096 --jinja --host 0.0.0.0 --port 8080

Recommended sampling (per the base model card): temperature between 0 and 0.2; avoid repetition penalties — they degrade instruction-following and tool-call validity.

Performance

Architecturally identical to ALIA-40b-instruct-2601 (LlamaForCausalLM, 48 layers, hidden 8192, 64/8 heads, vocab 256k). Expect equivalent throughput on Blackwell — ~10 tok/s generation on a single NVIDIA GB10 (DGX Spark) at -ngl 99 -c 4096 --jinja.

Backend	Format	Token gen (single prompt, GB10)
GPU (`-ngl 99`)	NVFP4	~10 tok/s
GPU (`-ngl 99`)	Q8_0	~5 tok/s

On non-Blackwell hardware, Q4_K_M / Q5_K_M / Q8_0 from the BF16 source remain better trade-offs because their kernels are more mature.

Quantization details

Source: BSC-LT/ALIA-40b-fc-2605 BF16 (~76 GB across 17 shards)
Tool: NVIDIA ModelOpt NVFP4 → convert_hf_to_gguf.py (llama.cpp master, post-PR #22611)
Scheme: NVFP4 (NVFP4_DEFAULT_CFG)
Calibration: 128 samples of HuggingFaceH4/ultrachat_200k (train_sft split), max_seq_len 2048
Ignored layers: lm_head

Calibration caveat

ultrachat_200k is English-only synthetic chat. The originally planned mix with Salesforce/xlam-function-calling-60k (function-calling traces) was blocked at run time because that dataset is gated on HF, so calibration fell back to the proven 2601 recipe. The model's FC distribution is already encoded in its SFT weights, but per-tensor scales were computed against a chat-only activation distribution, which may underweight tool-call activation patterns. BSC themselves note the FC fine-tune is "primarily evaluated and optimized for English," so English chat and tool calling should be well-served; multilingual tool calling (Spanish/Catalan/Basque/Galician) may be more affected by quantization noise than prose. Evaluate on your own multilingual tool-call task before deploying.

License & attribution

Released under the same Apache 2.0 license as the source.

Base model: BSC-LT/ALIA-40b-fc-2605 by Barcelona Supercomputing Center (BSC). Please cite their work if you use this model in research:

@misc{alia-40b-fc-2605,
  author = {Barcelona Supercomputing Center},
  title  = {ALIA-40b-fc-2605},
  year   = {2026},
  url    = {https://huggingface.co/BSC-LT/ALIA-40b-fc-2605}
}

NVFP4 quantization: Montevive AI.

Limitations

Inherits all limitations of the base ALIA-40b-fc-2605 model, including:

Not safety-aligned. BSC's model card explicitly notes the FC fine-tune is instruction-tuned but lacks value alignment. Apply your own guardrails before production use.
Tool calling is English-optimized. BSC's BFCL numbers (Non-Live Multiple AST 94.5%, Live Multiple AST 74.4%) are English; multilingual tool calling is on BSC's roadmap but not yet evaluated.
Multi-turn tool calling is weaker than single-turn (BFCL Multi-Turn Base 15.5% on BF16, per BSC's card). Don't expect long agentic loops without external scaffolding.

Plus standard NVFP4 quantization caveats:

NVFP4 inference quality is below Q8_0 on small models. ALIA-40b is well within NVFP4's design sweet spot (originally targeted at 70B+) but evaluate on your task before deploying.
CPU performance is poor — ARM NEON dot product for NVFP4 is much newer than for Q8_0 / K-quants. Use Blackwell GPUs for real workloads.

montevive/ALIA-40b-fc-2605-NVFP4-GGUF

ALIA-40b-fc-2605 — NVFP4 GGUF