Holo4-35B-A3B-exl3-4bpw

Source: Hcompany/Holo4-35B-A3B at exact revision 508458727831368da85d53a8055c2b0f923ed4f3. Full-precision BF16 source; source shard SHA-256s verified locally. Abliterated source: False. no verified abliterated source found

Original model by H Company; Apache-2.0 license included. Changes here: EXL3 quantization and removal of dangling MTP configuration.

Existing EXL3 converter revision 02aef45cd681b960a00afcd0749a4ab99e6c1bfe. Target 4 bpw; converter-reported bitrate 4.08; total tensor-storage bitrate 4.433904 bits per original source parameter, including scales and unquantized tensors. These measures have different denominators. HQ allocation, 8-bit output head, 6-bit vision linear weights, mul1 codebook, output scales always. Calibration: pinned converter bundled mixed corpus, 250 rows × 2048 tokens, deterministic seed 0; six corpus SHA-256s and exact quantization settings are in provenance.json. No measurement reuse.

Native tokenizer and native chat template retained; thinking defaults follow the original template. Validation used native template with thinking disabled. Source config declares MTP, but source tensor index contains no genuine MTP weights: its dangling layer count was set to zero in owned staging. No MTP draft model is included.

The actual exported language model was loaded with EXL3 and generated nonempty text before publication. This is a text smoke validation, not a benchmark of GUI grounding, vision, agent accuracy or refusal behavior. No original benchmark scores are claimed for this quantization. Use EXL3-compatible inference software; this is not an EXL2 or GPTQ checkpoint.

tensor-inventory.json, checksum-manifest.json and provenance.json describe every artifact and the validation. The manifest excludes its own file; its SHA-256 and remote publication commit are recorded in the campaign receipts.

Stop tokens and publication verification — 2026-10-12

The native tokenizer and chat template are retained. config.json sets eos_token_id: [248046, 248044] at the top level and in text_config, matching generation_config.json: <|im_end|> and <|endoftext|>. This alignment is checked from the configuration and tokenizer files.

Fresh local SHA-256 checks match the published artifact content. Safetensors header bounds and complete index-to-shard tensor coverage were checked for all 3 weight shards (125,233 indexed tensors). The checksum manifest covers the artifact files and excludes its own bytes.

MTP: this export contains zero MTP tensor keys in its safetensors headers and tensor index. Native MTP/speculative decoding is not included, and no MTP speed benchmark is claimed.

Earlier model-specific language smoke evidence is recorded in provenance.json. This publication verification did not rerun GPU generation or performance tests.