Synthia-4-27B-GGUF

Synthia-4-27B GGUF

Quantized releases of Synthia-4-27B, a personal AI and technical agent built on Qwen/Qwen3.8-27B. Synthia combines a consistent conversational personality with long-session continuity, tool use, multimodal input, and practical coding and engineering ability.

These GGUF builds are intended for running Synthia locally in a personal agent runtime. In testing, Synthia maintained its character and light humour across extended sessions while moving naturally between everyday conversation, planning, research, creative work, and multi-step technical tasks.

Every quantization is available with and without the inherited MTP next-token prediction layer.

Personal-agent setup

Provide Synthia's identity, the user's preferences, and any retrieved memories through the runtime's system context. The model can remain consistent with context supplied during a long session; persistence between separate sessions is managed by the host runtime.

Use the embedded Jinja chat template and pass tool definitions through the runtime's normal tool-schema interface. The template supports xhigh, medium, and low reasoning effort.

Main model files

llama.cpp does not define a Q8_K_M format; Q8_0 is the supported 8-bit release format.

Companion files

• Synthia-4-27B-mmproj-F16.gguf — Qwen vision projector/encoder companion.
• Synthia-4-27B-MTP-Q8_0.gguf — standalone MTP draft companion for runtimes that support split MTP deployment.

The Synthia LoRA trained language tensors only. Vision and MTP weights are the compatible, unchanged weights inherited from Qwen3.8-27B.

Recommended llama.cpp launch

The bundled-MTP files are recommended for current llama.cpp builds:

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 no-MTP launch, select a filename without -MTP and omit --spec-type draft-mtp.

The standalone MTP sidecar is structurally valid. llama.cpp build b10007-ef25ebc20 showed an M-RoPE position error on its separate-sidecar MTP path, while the bundled-MTP variants passed generation. Use a bundled variant with that build.

Validation

• All eight variants have matching tensor names and shapes within their family
• 851 tensors / 64 blocks without MTP
• 866 tensors / 65 blocks with one MTP layer
• Embedded Qwen3.8 tool/reasoning chat template preserved in every variant
• Vision request succeeded through the F16 projector
• Bundled MTP drafted 66 tokens and accepted 46 in the multimodal smoke test
• SHA-256 hashes are provided in SHA256SUMS