Tdamre/limite-1b-base-GGUF-LiteRT

limite-1b-base: GGUF and LiteRT-LM

Community conversions by Tdamre of Paradigma's original model, under Apache 2.0. Source revision: c55f6dd9741d89b88186235c6d43bff367ae3cb0. Both formats preserve the model's custom Limite architecture.

GGUF runtime requirement: these files require the included Limite llama.cpp patch. Stock llama.cpp/Ollama builds do not implement this architecture at this release. The LiteRT-LM bundles use the standard LiteRT-LM format and were tested with LiteRT-LM 0.17.0 on CPU. Use the CPU backend for these graphs; the tested GPU delegate rejects operations in them.

This is the base pretrained checkpoint. Use raw text completion or few-shot prompts; it is not instruction-tuned. The upstream tokenizer/template files are preserved for provenance.

GGUF files

Each GGUF is usable with 16,384, 32,768 or 65,536 token runtime contexts. The original 131,072-position metadata is retained; separate copies of identical weights are unnecessary. Presets specify the requested contexts.

File	Size (GiB)	Short-prompt KL vs source FP32	Top-token agreement
Q2_K_L	0.457	0.29995	87.5%
Q3_K_M	0.592	0.03134	100.0%
Q4_K_M	0.684	0.01729	100.0%
Q5_K_M	0.769	0.01078	100.0%
Q6_K	0.859	0.01103	100.0%
Q8_0	1.093	0.00171	100.0%

These are public Unsloth exporter + model-specific importance-matrix quantizations. The exact private Unsloth Dynamic v2/v3 generator was unavailable, so these are not branded as Dynamic v2/v3 or official Unsloth releases. The lowest-bit files trade substantial accuracy for size; Q4_K_M through Q8_0 offer progressively more weight precision.

LiteRT-LM files

Mixed INT4/INT8 precision: Q/K/V and feed-forward gate/up projections use OCTAV INT4/block32; attention-output/down projections and vocabulary tables use INT8 with floating computation. Small gates remain float. Graph interfaces and KV caches are float32; INT4 projections use dynamic integer computation internally. Each file contains 519,045,120 physically stored INT4 weight parameters and 516,096,000 INT8 weight parameters, independently audited after export. Each .litertlm contains its TFLite graph, original tokenizer, original chat template and stop-token metadata.

Compiled context	Bundle size (GiB)	KV tensors alone
16,384 tokens	0.783	1.5 GiB
32,768 tokens	0.798	3.0 GiB
65,536 tokens	0.830	6.0 GiB

Weights, temporary buffers, delegate preparation and other runtime allocations require additional RAM. The 64K bundle is intended for devices with sufficient memory. GPU/NPU execution and physical Android/iOS devices have not been certified in this release.

For long CPU generations, use the CLI's streaming path or the Python asynchronous APIs. LiteRT-LM 0.17.0 synchronous decode has a ten-minute deadline. This release prioritizes numerical fidelity; CPU setup and inference at large contexts can be slow.

Run

Build the GGUF runtime using runtime/README.md. For raw completion:

tools/llama.cpp/build/bin/llama-completion \
  -m gguf/limite-1b-base-Q4_K_M.gguf -c 16384 -ngl 99 \
  --no-context-shift -n 256 -p "The capital of France is"

For Violetto chat, use llama-server with the embedded Jinja template and user-only messages; preserve its fixed mathematical system prompt. For the base checkpoint, use completion prompts.

LiteRT-LM example (choose 32768 or 65536 in both places for the larger bundles):

litert-lm run litert/limite-1b-base-int4-ctx16384.litertlm \
  --backend cpu --max-num-tokens 16384 --cache disk \
  --temperature 0.6 --top-p 1.0 --top-k 50 \
  --prompt "What is 2 + 2?"

For the base model, add --no-template and use a text-completion prompt. Its bundle stops on both the tokenizer's endoftext and generation-config im_end tokens. Violetto's original generation config specifies temperature 0.6. Verification used greedy sampling for repeatability. Reasoning is text generated by the model; this bundle does not advertise an on/off thinking control.

Conversion and checks

There was no retraining. Each checkpoint was separately calibrated on 327,680 tokens drawn from a deterministic GSM8K-training/MBPP corpus. The importance matrix and reproduction scripts are included. Sensitive Limite gates, dense residual mixers and rotary tables are preserved. The runtime handles large feed-forward activations without clipping the trained function.

Checks cover all six GGUFs at all three contexts, finite logits at the final position, numerical comparisons, cached generation, compiled LiteRT cache shapes, and exact packaged tokenizer/template/graph identity. validation.json contains the evidence; validation_inputs.json lists the fixed inputs. The GGUF table uses one short prompt. LiteRT comparison uses 43 positions in Violetto's formatted prompt or 48 positions pooled across five raw base-model prompts, with the same KL and agreement thresholds. These are smoke comparisons, not a benchmark suite. No upstream math-benchmark score is claimed for these quantizations, and full-length retrieval quality was not evaluated.

File hashes are in SHA256SUMS; exact file sizes, source pin and format details are in artifact_manifest.json.

License and attribution

Apache License 2.0, carried over from Paradigma's release. See NOTICE for attribution and tool licenses. Paradigma's model card and release announcement describe the original models. These community conversions are not endorsed releases from Paradigma, Unsloth, Google or llama.cpp.