Synthia-4-27B

Synthia-4-27B is a personal AI with the ability to do real work. It combines an expressive, conversational presence with the tool use and persistence needed for coding, research, planning, creative work, and day-to-day assistance.

Synthia has character. It can be warm, candid, and lightly funny without turning every exchange into a performance. More importantly, it can hold that voice across a long session while moving naturally between conversation and execution.

The model starts from Qwen/Qwen3.8-27B and is post-trained on complete, long-form agent sessions at a 65,536-token training length. The training objective covers every assistant turn, including tool calls, so Synthia learns how a working relationship develops across a task rather than only how to produce an isolated answer.

A personal AI that can act

Synthia has been tested in a personal AI agent runtime, where it showed strong continuity over extended sessions. It maintained a recognizable personality, remembered the active conversational context, used humour appropriately, and remained oriented while working through multi-step tasks with tools.

Its intended role is broader than a coding assistant. Synthia can discuss an idea, help make a decision, organize a project, work through a difficult technical problem, or simply be good company while doing all of the above. When the runtime supplies durable memories or personal context, Synthia can incorporate them into the current conversation; persistence between separate sessions remains the responsibility of the host runtime.

Behavior profile

Synthia is tuned to:

• maintain a stable voice and relationship with the user across long sessions;
• balance personality and light humour with direct, useful answers;
• move smoothly between open conversation and task execution;
• preserve goals and constraints across extended tool-driven work;
• inspect the available evidence before committing to a solution;
• revise a plan when tool results contradict an earlier assumption;
• treat implementation and verification as parts of the same task;
• express uncertainty when the available evidence does not support a firm claim; and
• vary its reasoning budget through the bundled chat template.

It retains the base model's image and video input path, 262,144-token native context window, and multi-token prediction (MTP) head. Post-training examples were limited to 65,536 tokens, so behavior beyond that length comes from the base model rather than from the fine-tuning distribution.

Prompting

Use the tokenizer and chat template shipped in this repository. A personal-agent runtime should provide Synthia's identity, the user's preferences, and any retrieved memories in the system context. The template supports xhigh, medium, and low reasoning effort and formats reasoning inside <think>...</think> blocks.

prompt = processor.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    reasoning_effort="xhigh",
)

Pass tool definitions with the template's tools= argument. Synthia was trained on conversations containing system instructions, user requests, assistant messages, tool calls, and tool results.

Files and companion releases

This repository contains the merged BF16 Transformers checkpoint.

Format	Approximate size	Typical use
BF16 safetensors	55.6 GB	Transformers, vLLM, SGLang, conversion

Quantized builds are available in migtissera/Synthia-4-27B-GGUF.

Quantization	Standard	MTP bundled
F16	50.11 GiB	50.90 GiB
Q8_0	26.63 GiB	27.05 GiB
Q6_K	20.57 GiB	20.89 GiB
Q4_K_M	15.41 GiB	15.66 GiB

The GGUF repository also provides an F16 vision projector and a standalone Q8_0 MTP companion file. Use a bundled-MTP model with a compatible llama.cpp build when speculative decoding is desired. Standard GGUF files are available for runtimes without MTP support.

Transformers example

Install a Transformers version that supports Qwen3.8, then load the processor and model from this repository:

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "migtissera/Synthia-4-27B"

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {
        "role": "system",
        "content": "You are Synthia, my personal AI. Be candid, capable, warm, and concise. Use tools when they help you complete the work.",
    },
    {
        "role": "user",
        "content": "Help me choose what to focus on today, then inspect the project and get the first task moving.",
    },
]

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    reasoning_effort="xhigh",
    return_tensors="pt",
).to(model.device)

output = model.generate(inputs, max_new_tokens=2048)
print(processor.decode(output[0], skip_special_tokens=True))

llama.cpp example

Download a bundled-MTP Q4_K_M build and the vision projector:

hf download migtissera/Synthia-4-27B-GGUF \
  Synthia-4-27B-Q4_K_M-MTP.gguf \
  Synthia-4-27B-mmproj-F16.gguf \
  --local-dir ./synthia-4-27b

Start the server:

llama-server \
  --model Synthia-4-27B-Q4_K_M-MTP.gguf \
  --mmproj Synthia-4-27B-mmproj-F16.gguf \
  --spec-type draft-mtp \
  --ctx-size 65536 \
  --parallel 1 \
  --gpu-layers 99 \
  --flash-attn auto \
  --jinja \
  --image-min-tokens 1024

For a standard GGUF, choose a filename without -MTP and remove --spec-type draft-mtp.

Where Synthia fits

• A persistent personal AI in a stateful agent runtime
• Daily planning, decision support, writing, and creative collaboration
• Long-running research and technical work with many observations
• Repository exploration, implementation, debugging, and verification
• Tool-driven workflows with structured function definitions
• Image-assisted conversation and analysis

Training record

Setting	Value
Training data	Curated long-form agentic sessions
Sequence length	65,536 tokens
Epochs / optimizer steps	2 / 30
Batch size	8
LoRA rank / alpha	32 / 32
Learning rate	1e-4, linear decay
Supervised tokens	All assistant messages and tool-call turns
Validation NLL	0.73384 → 0.67021

The adapter targeted the language model. The vision encoder, projector, and MTP head were inherited unchanged from the base checkpoint. The published BF16 weights already include the language-model adapter and do not require a separate LoRA at inference time.

Artifact checks

The release was checked for the following properties:

• 1,199 BF16 tensors are present across 18 safetensor shards;
• tensor names and shapes agree with the base checkpoint;
• trained language projections differ from the base while untargeted embeddings remain identical;
• the tokenizer and chat template are preserved;
• text generation, image input, and bundled-MTP decoding run with llama.cpp on Apple Metal; and
• file hashes are listed in SHA256SUMS.

Base model and license

Synthia-4-27B is derived from Qwen/Qwen3.8-27B. The architecture, tokenizer, multimodal stack, long-context support, and MTP components originate with the Qwen team.

The model is released under the Apache License 2.0. See LICENSE.

Citation

@misc{tissera2026synthia4,
  title        = {Synthia-4-27B},
  author       = {Migel Tissera},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/migtissera/Synthia-4-27B}},
  note         = {A multimodal personal and technical agent fine-tune of Qwen3.8-27B}
}