Brought to you by @ContextReq
Pebble-25M-Chat-GGUF
GGUF conversions of basically-ai/Pebble-25M-Chat (Apache 2.0).
IMPORTANT: patched llama.cpp required
Pebble uses a custom hybrid Mamba2 + attention architecture. These GGUFs carry
general.architecture = "pebble", which upstream llama.cpp refuses to load.
Everything needed to run them lives in the support repo:
rootendpoint/basicallyai_llama.cpp_support
llama.cpp-pebble.patch- adds thepebblearchitecture to llama.cpp (applies cleanly against upstream commit0eadefe)basicallyai_to_gguf.py- standalone converter (numpy + safetensors only)numpy_reference.py- independent reference implementation used to verify correctness
The chat template is embedded in the GGUF metadata; chat mode works out of the box:
llama-cli -m pebble-25m-chat-f16.gguf
Chat format: lowercase user: ... / assistant: ... turns.
Files
| Quant | Size | Type |
|---|---|---|
| f16 | 49.1 MB | F16 |
| q8_0 | 26.3 MB | mostly Q8_0 |
| q4_k_m | 18.5 MB | mostly Q4_K_M |
Verification
Outputs were cross-checked token-by-token against an independent pure-numpy reference implementation over an 8-prompt battery (CPU and CUDA backends). Identical greedy sequences on all prompts. Quantized builds verified greedy-identical to f16 (8/8 prompts for both q8_0 and q4_k_m).
Model
- 25M parameters, hidden 608, 8 layers (6 Mamba2 + 2 attention), ctx 2048, vocab 2048
- SFT chat variant; a research-scale model: expect toy-level output quality.
CPU support (no GPU required)
Pure-CPU support for these models (no mamba-ssm, no CUDA) lives in the basicallyai_cpu_support repository:
https://github.com/rootendpoint/basicallyai_cpu_support
It runs the original HF checkpoints in pure PyTorch on plain CPU (~118 tok/s for 10M, ~65 tok/s for 25M on a Ryzen 5 2600X), verified token-identical against this GGUF pipeline and an independent numpy oracle.