talkie-1930-13b-it-q5

Talkie 1930 13B IT Q5

talkie-1930-13b-it-q5 is an experimental Q5_0 GGUF build derived from talkie-lm/talkie-1930-13b-it.

This is a consumer-GPU-oriented artifact for running Talkie locally through a patched llama.cpp runtime. It is not yet compatible with stock llama.cpp or stock LM Studio because Talkie uses a custom architecture.

File

talkie-1930-13b-it-q5.gguf

Expected SHA256:

B6025276018B228CB35CDC76F2D957EB8037EA595F7C9486BC2971ECAFEAD0BA

Runtime

Use the companion GitHub repo for the Node app, patched llama.cpp runtime, build scripts, and release notes:

https://github.com/solwyc/talkie-1930-13b-it-q5

Tested local Windows settings:

GPU:          RTX 4070 12 GB
GGUF:         Q5_0, about 9.13 GB
offload:      -ngl 26
context:      1024
speed:        roughly 7-8 tok/s in short local tests

The repo includes experimental Linux CUDA and macOS Metal build scaffolding, but those ports were drafted blindly with GPT-5.5 assistance and have not been maintainer-tested yet.

Important Compatibility Note

This GGUF has:

general.architecture = talkie

It requires a Talkie-aware llama.cpp runtime with support for Talkie's custom blocks, tensor mapping, QK norm/gain tensors, and RMSNorm epsilon behavior.

Build Summary

The local build path was:

1. Start from Talkie 1930 13B IT.
2. Train and select a small local LoRA polish adapter.
3. Merge the adapter into the BF16 checkpoint.
4. Convert the merged checkpoint to a Talkie GGUF.
5. Quantize the GGUF to Q5_0.
6. Validate through patched llama.cpp with CPU/GPU offload.

Limitations

• Experimental runtime patch required.
• Not a stock LM Studio model yet.
• May produce plausible but wrong historical claims.
• May invent personal identity details unless prompted carefully.
• Not for high-stakes use.

Attribution

Based on the Talkie 1930 13B IT model and code from:

https://github.com/talkie-lm/talkie

Upstream Talkie authors named in the project README: Alec Radford, Nick Levine, and David Duvenaud.