Capability matrix
- Compute and format Q5_K_M · Commodity x86 nightly lanes.
- Local entry point Text REPL.
- Server Local HTTP pass. One-word reply; 1K build.
- Harness Tool workflow unverified. Run a named client against an identified bundle.
Choose the download
Publisher model card · CKE input files
Publisher task: text generation. CKE example: Q5_K_M GGUF. Use this artifact identifier:
hf://unsloth/gemma-3-270m-it-GGUF/gemma-3-270m-it-Q5_K_M.gguf.
Get started
--chat-template auto
version/v8/scripts/cks-v8-run run \
hf://unsloth/gemma-3-270m-it-GGUF/gemma-3-270m-it-Q5_K_M.gguf \
--context-len 1024 --force-compile --force-convert --chat-template=auto \
--prompt "Give me a concise example of C code." \
--max-tokens 64 --temperature 0.0 \
--generate-visualizerAn earlier cached Q5_K_M trial returned an empty answer. A fresh 1K build from the same GGUF answered in both the terminal and the HTTP trial. The exact comparisons below were recorded on different Gemma3 quantizations.
What has actually run?
256/256 complete-vocabulary rows bit-exact to llama.cpp for both artifacts via a circuit-declared numerical schedule. 520-token boundary nearly exact but differs at five rows; longer-context bit-exact certification open.
CPU evidence. Commodity x86 nightly lanes; schedule contract documented in the numerical-contracts case study.
Performance. The cited model result does not give a comparable prefill/decode pair here. ARM throughput for this exact artifact is unverified here. Benchmark evidence · 32K to 128K results
Can I fine-tune it?
Full-checkpoint training, LoRA and other PEFT methods: no verified CKE recipe for this exact checkpoint in the reviewed evidence. For a trainable starter, use the FP32 and adapter recipes.
Start an interactive session
Run from the CKE checkout. Type a question, then /exit to finish.
version/v8/scripts/cks-v8-run run \
hf://unsloth/gemma-3-270m-it-GGUF/gemma-3-270m-it-Q5_K_M.gguf \
--chat-template auto --max-tokens 128 --temperature 0