SAVRN
Search Contact SAVRN

Open-weight model · Text generation

mato-v1.0-sft

by German Gimenez german-gimenez/mato-v1.0-sft

Mato es el asistente de IA entrenado en casa por Napsix — una plataforma B2B agéntica donde los agentes corren solos: investigan, escriben, operan herramientas y producen resultados (documentos, planillas, tableros, formularios, sitios).

Parameters
Context
Weights934.0 MB
Licenseapache-2.0
AccessAccess requested at publisher
Monthly Downloads82

Model Card

By German Gimenez, published under apache-2.0, revision d589b9efcb9c.

Mato es el asistente de IA entrenado en casa por Napsix — una plataforma B2B agéntica donde los agentes corren solos: investigan, escriben, operan herramientas y producen resultados (documentos, planillas, tableros, formularios, sitios). Este repositorio contiene el adapter LoRA (SFT) sobre Qwen/Qwen3.8-27B. En producción se sirve como LoRA sobre el checkpoint FP8 del base, con MTP speculative decoding y 256k de contexto, en GPUs spot de Daytona. - Tool calling nativo en el formato XML de Qwen (parser qwen3xml de vLLM): llamadas correctas, argumentos válidos y recuperación ante fallos. directo y conciso, sin relleno ni muletillas. - Conocimiento de las superficies de Napsix: XIA, Agentes…

Read German Gimenez's full model card

Mato V1.0 — el modelo insignia de Napsix

Mato es el asistente de IA entrenado en casa por Napsix — una plataforma B2B agéntica donde los agentes corren solos: investigan, escriben, operan herramientas y producen resultados (documentos, planillas, tableros, formularios, sitios).

Este repositorio contiene el adapter LoRA (SFT) sobre Qwen/Qwen3.8-27B. En producción se sirve como LoRA sobre el checkpoint FP8 del base, con MTP speculative decoding y 256k de contexto, en GPUs spot de Daytona.

[!TIP] ¿Querés probarlo ya, gratis y sin clonar nada? Mato está incluido en el AI Gateway de Napsix (API OpenAI-compatible): https://www.napsix.com/api/gateway/v1 · modelo napsix/mato-v1.0Crear una API key · Quickstart en 2 minutos

Qué puede hacer

  • Tool calling nativo en el formato XML de Qwen (parser qwen3_xml de vLLM): llamadas correctas, argumentos válidos y recuperación ante fallos.
  • Voz XIA: español neutro profesional (también portugués e inglés), directo y conciso, sin relleno ni muletillas.
  • Conocimiento de las superficies de Napsix: XIA, Agentes, Espacios (Rooms, Projects, Inbox, CRM, Napsix Mail), Resultados (Docs, Sheets, Dashboards, Forms, Sitios) y Conexiones (1000+ apps, WhatsApp, MCP, APIs).
  • Conducta de producto: no inventa datos; pide clarificación (ask_clarification) cuando la tarea es ambigua; escala en vez de prometer lo que no puede.
  • Heredado del base Qwen3.8-27B: contexto 256k, visión y multilingüismo.

La historia (por qué existe Mato)

Napsix cree en eficiencia, no en fuerza bruta: cada tarea va al modelo del tamaño correcto. Pero el asistente de la casa necesitaba tres cosas que ningún modelo externo daba juntas: control del protocolo de herramientas, privacidad de los datos del producto y costo predecible. Entonces lo entrenamos:

  1. F0–F1 · Dataset: trazas reales de la plataforma (anonimizadas) + destilación con un teacher in-family (qwen3.8-max) → conversaciones con el chat template real de Qwen y tool-calls verificables.
  2. F2 · Entrenamiento: QLoRA (NF4, r=32, all-linear) en 1×H100 spot, con checkpoints y resume real ante evictions.
  3. F3 · Evaluación: suite de tool-calling + judge LLM — mejora clara de comportamiento sin regresión en herramientas.
  4. F4 · Deploy: LoRA servido en producción (FP8 + MTP + 256k) con el mismo slug napsix/mato-v1.0 y rollback de un comando.
  5. Loop: el corpus crece con la plataforma; un pipeline semanal propone re-entrenamientos con aprobación humana.

Mato V1.0 está sirviendo en producción desde el 2026-09-17.

Cómo usarlo

Opción A — API (recomendada, gratis)

curl https://www.napsix.com/api/gateway/v1/chat/completions \
  -H "Authorization: Bearer nsx_live_TU_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"napsix/mato-v1.0","messages":[{"role":"user","content":"Hola Mato"}]}'

También es drop-in con el SDK de OpenAI: cambiá baseURL y listo (docs).

Opción B — local (transformers + peft)

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3.8-27B", torch_dtype="bfloat16", device_map="auto"
)
tok = AutoTokenizer.from_pretrained("german-gimenez/mato-v1.0-sft")
model = PeftModel.from_pretrained(base, "german-gimenez/mato-v1.0-sft")

Opción C — vLLM (producción)

vllm serve Qwen/Qwen3.8-27B-FP8 \
  --enable-lora --max-lora-rank 32 \
  --lora-modules mato-tuned=german-gimenez/mato-v1.0-sft \
  --tool-call-parser qwen3_xml --enable-auto-tool-choice

Ficha técnica

Base Qwen/Qwen3.8-27B (Apache-2.0)
Adapter LoRA all-linear · r=32 · alpha=64 · 233M params (0.86%)
Entrenamiento QLoRA NF4, seq 4096, 6 epochs, 1×H100 spot
Eval (judge LLM) comportamiento: 6–2–1 vs base · tool-calling: sin regresión
Serving FP8 + LoRA + MTP + 256k · 3 réplicas spot auto-gestionadas
Idiomas Español (neutro B2B) · Portugués · Inglés

Links

  • Mato — https://www.napsix.com/mato
  • Gateway API (gratis) — https://www.napsix.com/gateway
  • Tu API key — https://www.napsix.com/dashboard/ai-gateway
  • Developers — https://www.napsix.com/developers
  • Status — https://www.napsix.com/status

Uso y licencia

  • Licencia: Apache-2.0 (heredada del base). Atribución a Napsix y a Qwen Team apreciada.
  • Mato está pensado para asistentes B2B legítimos. No usar para spam, desinformación, suplantación de identidad ni contenido dañino.
  • "Mato" es una marca de Napsix. Si construís sobre Mato, aclará que es un modelo de Napsix (ej. "powered by Mato V1.0 — Napsix").

About (English)

Mato V1.0 is Napsix's in-house assistant model: a LoRA SFT adapter over Qwen/Qwen3.8-27B, trained on real (anonymized) product traces distilled with an in-family teacher. It excels at tool calling, the XIA brand voice (neutral professional Spanish, Portuguese, English) and Napsix product knowledge — grounded, concise and honest. Serving: FP8 + LoRA + MTP + 256k on spot H100s. Try it free via our OpenAI-compatible gateway: https://www.napsix.com/api/gateway/v1 (model napsix/mato-v1.0) — https://www.napsix.com/mato

Mato V1.0 by Napsix — https://www.napsix.com/mato

Identity and Version

Repository
german-gimenez/mato-v1.0-sft
Publisher
German Gimenez
Task
Text generation
Modality
Text
Library
peft
Parameters
Not stated by the source
Languages
es, pt, en
Revision
d589b9efcb9cf4a2c3ce47239109f1b4879e9f71
First published
2026-09-17
Last updated
2026-09-18

Files and Weights

8 files, 954.0 MB in total. The weights are 2 files totalling 934.0 MB in bin, safetensors.

Weights2 files · 934.0 MB
Configuration1 file · 1.3 KB
Tokenizer2 files · 20.0 MB
Documentation1 file · 6.3 KB
Other1 file · 9.0 KB
Repository1 file · 1.6 KB
Every file
FileTypeSizeSHA-256
adapter_model.safetensorsWeights934.0 MB
training_args.binWeights5.2 KB
adapter_config.jsonConfiguration1.3 KB
README.mdDocumentation6.3 KB
chat_template.jinjaOther9.0 KB
.gitattributesRepository1.6 KB
tokenizer.jsonTokenizer20.0 MB
tokenizer_config.jsonTokenizer1.1 KB

License and Download

License
apache-2.0
Access
Access requested at publisher
Download size
934.0 MB
Request access from German Gimenez

German Gimenez grants access through its official repository on Hugging Face. Read the license.

Built From

Memory Requirements

PrecisionWeights in memory
As published934.0 MB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Questions About mato-v1.0-sft

Can I use mato-v1.0-sft commercially?

Yes. mato-v1.0-sft is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

Similar Models

Fine-tune Qwen3 (14B) for free using our Google Colab notebook! - Read our Blog about Qwen3 support: unsloth.ai/blog/qwen3 - View the rest of our notebooks in our docs here. Qwen3-Coder is available in multiple sizes. Today, we're excited to introduce Qwen3-Coder-30B-A3B-Instruct. This streamlined model maintains impressive performance and efficiency, featuring the following key enhancements: - Significant Performance among open models on Agentic Coding, Agentic Browser-Use, and other foundational coding tasks. - Long-context Capabilities with native support for 256K tokens, extendable up to 1M tokens using Yarn, optimized for repository-scale understanding. - Agentic Coding supporting for…

Open weights apache-2.0 transformers

Model · Text generation

opt-125m

AI at Meta

OPT was first introduced in Open Pre-trained Transformer Language Models and first released in metaseq's repository on May 3rd 2022 by Meta AI. Disclaimer: The team releasing OPT wrote an official model card, which is available in Appendix D of the paper. Content from this model card has been written by the Hugging Face team. To quote the first two paragraphs of the official paper OPT was predominantly pretrained with English text, but a small amount of non-English data is still present within the training corpus via CommonCrawl. The model was pretrained using a causal language modeling (CLM) objective. OPT belongs to the same family of decoder-only models like GPT-3. As such, it was…

Open weights other 2,048 tokens transformers

Model · Text generation

Ornith-1.5-9B-GGUF

Ornith

Chirp Chirp! We are introducing Ornith-1.5, a major step toward building foundation models through end-to-end self-improvement. Ornith-1.5 extends Ornith-1.0 (which was developed on top of Qwen3.5 and Gemma4 with additional continued pretraining, mid-training, and post-training) by expanding the self-improvement loop from scaffold and rollout optimization to jointly optimizing task generation, scaffold construction, and solution rollouts. Rather than relying on a fixed set of human-curated tasks and manually designed harnesses, Ornith-1.5 continuously generates new training tasks, discovers effective strategies for solving them, and improves the policy through reinforcement learning. For…

Open weights mit transformers

Model · Text generation

Ornith-1.5-35B-A3B-GGUF

Ornith

Chirp Chirp! We are introducing Ornith-1.5, a major step toward building foundation models through end-to-end self-improvement. Ornith-1.5 extends Ornith-1.0 (which was developed on top of Qwen3.5 and Gemma4 with additional continued pretraining, mid-training, and post-training) by expanding the self-improvement loop from scaffold and rollout optimization to jointly optimizing task generation, scaffold construction, and solution rollouts. Rather than relying on a fixed set of human-curated tasks and manually designed harnesses, Ornith-1.5 continuously generates new training tasks, discovers effective strategies for solving them, and improves the policy through reinforcement learning. For…

Open weights mit transformers

Model · Text generation

Ornith-1.0-9B-GGUF

Ornith

Aloha! Today, we are releasing Ornith-1.0, a self-improving family of open-source models for agentic coding. This model card documents Ornith-1.0-9B, the most lightweight member of the Ornith family, designed for efficient single-GPU deployment. Ornith-1.0-9B is a dense ~9B model (≈19 GB in bf16), so it serves comfortably on a single 80GB GPU. The recipes below stand up an OpenAI-compatible server; add --tensor-parallel-size / --tp if you want to shard across more GPUs. For a quick local test (or to script offline generation), load the model directly with Transformers. Make sure you have a recent release installed — see the Transformers installation guide; Ornith-1.0-9B requires…

Open weights mit transformers

Uncensored Qwen3.8-27B, published as GGUF quantizations with the multi token prediction (MTP) head retained and verified. Refusal behaviour has been substantially reduced, not eliminated. See Measured behaviour for the numbers. Capabilities, training data, and architecture are otherwise unchanged. - Refusal directions removed with Heretic, which co minimizes refusal count against KL divergence from the base model. No handwritten refusal removal code, no finetuning, no additional training data. - Abliteration runs at bf16 (no 4 bit quantization). the resulting LoRA is merged into the bf16 base, so the published weights are not a quantized round trip. - mtp. tensors are copied verbatim from…

Open weights apache-2.0 llama.cpp