Pluto-AI-Labs / Athena-1.0-4B-GGUF

Athena 1.0 4B — GGUF

GGUF distribution of Athena 1.0 , developed by Pluto AI Labs through supervised adaptation of Qwen3.5-4B. The source is a standalone BF16 model with its trained language, vision and merger LoRA updates already merged. No separate PEFT adapter is required.

Release status: experimental. Quantization reduces storage and memory; it adds no training and establishes no benchmark improvement. GGUF exports receive structural and short runtime checks, not benchmark-quality certification.

Model and training

The final training checkpoint was selected before evaluation. The 81.57% monitoring-NLL reduction is not an 81.57% accuracy gain. Native context configuration is inherited, but long-context quality was not established by this training. There was no dedicated video training block.

Downloads

Pending language quantizations: IQ4_NL, Q3_K_M, Q3_K_L, Q3_K_S, IQ3_M, IQ3_S, IQ3_XS, IQ3_XXS, Q2_K, Q2_K_S, Q2_0, IQ2_M, IQ2_S, IQ2_XS, IQ2_XXS, IQ1_M, IQ1_S, Q1_0, TQ2_0, TQ1_0.

Q5_K_M/Q4_K_M are practical starting choices; Q6_K/Q8_0 or F16 favor fidelity over size. These are general quantization choices, not an Athena-specific measured ranking. All quantizations are made directly from F16, never by requantizing another integer file. Norms, sensitive tensors retained by upstream rules, and tensors lacking calibration coverage can remain at higher precision. The manifest reports actual tensor-type distributions and file averages.

Q1_0 has nominal 1.125 bits per weight, IQ1_S about 1.56 and IQ1_M about 1.75. None is an exactly one-bit total model. IQ1/Q1, 2-bit and ternary TQ formats are experimental compression options with potentially substantial quality loss. The integer formats' names describe upstream mixtures, not uniform precision for every tensor. Speculative MTP heads are excluded from this standard decoder export. Duplicate aliases, the MoE-only MXFP4_MOE format and redundant BF16/F32 copies are omitted.

Calibration uses deterministic text-only examples from the recorded Athena training-data revision, excluding identity rows: 16 chunks of 512 tokens. This is bounded CPU calibration, not a high-coverage multimodal calibration study. Original calibration text and benchmark media are not redistributed. Example keys, text checksum, actual chunks and tensor coverage are recorded in calibration/recipe.json and export_manifest.json .

Image inference

Download one language GGUF and mmproj-Athena-1.0-4B-F16.gguf . The shared F16 vision encoder/projector preserves the merged vision weights across language quantizations. The language file alone supports text inference. Image use requires a client supporting Qwen3.5 and its projector; generic GGUF support alone is insufficient. A small-image runtime probe does not certify every client's video support.

Build llama.cpp at the recorded revision or use a compatible later version. With both files downloaded:

./build/bin/llama-mtmd-cli -m Athena-1.0-4B-Q5_K_M.gguf \
  --mmproj mmproj-Athena-1.0-4B-F16.gguf \
  --image example.png -p "Describe this image." \
  --jinja -n 256

Deployment system prompt: You are Athena 1.0, a vision-language model by Pluto AI Labs. The native chat template is retained. Trained identity is a behavioral bias, not an immutable security boundary.

Existing source-model evaluation

These results concern the BF16 Transformers source, not GGUF inference. Both Athena and Qwen3.5-4B used the same selected samples with non-thinking greedy decoding and bounded outputs. This was not a reproduction of Qwen's published protocol. Values are percentages.

Coverage is 209 pairs, not full datasets. IFBench had 1 Athena and 2 base cap failures; strict accuracy was 23.81% for both. HMMT had 3/5 Athena and 5/5 base cap failures, limiting interpretation. One Athena document timed out. Video covered three videos and lacked subtitles for three questions. The video base score includes a disclosed uniform CPU correction of saved final-option extraction from 22.22% to 33.33%; no inference retry was used. ERQA/video leads each amount to one correct question. Overall superiority over the base or 7B/9B models is not established.

Audit archive SHA256: a80edbdd8831a3fb6f6eedf86a91d35def6a7d289ebac1c5865949f724beda6c . Protocol: fdc8db24041e708a9ca193cdd2cb275d57f35f9d7e2567fe107a08f79eb35ee2 . Base revision: 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a . The original evaluation's completion/quality gates did not all pass. Execution completion is not release-quality certification.

Intended use and limitations

Research and application development involving documents, visual questions, charts, OCR and conversation. Outputs can hallucinate, misread text or produce incorrect reasoning. The bounded source evaluation showed notable multilingual and document-extraction regressions. Identity and JSON formatting remain imperfect. Evaluate the chosen quantization on representative deployment inputs. Runtime memory also includes context, recurrent state and the vision projector.

License and provenance

Weights inherit Apache 2.0; the source LICENSE is included. The training dataset combines upstream sources with separate terms; consult its card and original collections. This statement does not relicense underlying data.

export_manifest.json records source/tool revisions, file SHA256 hashes and sizes, runtime probes, calibration coverage and failures. Each successful file is uploaded with the updated manifest/card and verified against its Hub LFS hash before local cleanup. Rerunning the same cell skips matching cloud-verified exports. Incomplete formats are not advertised as successful. The source Transformers repository is unchanged.