Groot-19B-A3B-261023 APEX-I MiniPlus V2.1

APEX Quantization Series by IsValorum

[!IMPORTANT] Independently computed using dedicated cloud compute clusters, funded 100% out-of-pocket.

If you consider these releases to be of high quality and they help you in your day-to-day workflow, consider supporting on Ko-fi. Every contribution directly sustains future cloud compute hours and helps support my university studies.

Family cross-links

Only the MiniPlus V2.1 tier is published for this model. NanoPlus and MicroPlus are not part of this release.

Architecture and precision

• Base model: GrootTeam/Groot-19B-A3B-261023 (Apache-2.0)
• Architecture: MoE with 40 language layers plus one MTP block; 30 linear-attention/SSM layers and 10 full-attention layers; 128 routed experts plus 1 shared expert, 8 routed experts active per token; hidden size 2048; expert intermediate size 512; 262,144-token context.
• iMatrix: public matrix from bartowski, GrootTeam_Groot-19B-A3B-261023-imatrix.gguf; 573 chunks of 512 tokens from calibration-v6; 510 entries. It covers 400 of the 413 quantized tensors. The 13 uncovered tensors are output.weight, token_embd.weight and the MTP block, which use K-quants or Q4_0 and do not need an importance matrix.
• Recurrent shielding: ssm_alpha, ssm_a, ssm_dt.bias, ssm_conv1d and all RMSNorms remain F32.
• Vision: two companion projectors are available: mmproj-GrootTeam_Groot-19B-A3B-261023-f16.gguf (original F16) and mmproj-GrootTeam_Groot-19B-A3B-261023-q8_0.gguf (Q8_0, 610.7 MB). Vision quality was not re-evaluated after quantization.
• MTP: block 40 is included in the main GGUF. Use --spec-type draft-mtp with llama.cpp b12017 or newer.
• Tensor count: 753 tensors, matching the 753-rule MiniPlus V2.1 MTP recipe.

File and metrics

Audited tensor map sample

Full tensor map

Complete list: 753 tensors

The machine-readable version is tensor-map.tsv.

Files

Usage

Tested with llama.cpp b12017. Groot needs a recent build; older builds such as b11512 may not load this model or its MTP block.

llama-cli

llama-cli \
  -m Groot-19B-A3B-261023.APEX-I-MiniPlus-V2.1.gguf \
  -p "Explain why the sky is blue in one paragraph." \
  -ngl 99 -c 32768 -t 8

llama-server

llama-server \
  -m Groot-19B-A3B-261023.APEX-I-MiniPlus-V2.1.gguf \
  --mmproj mmproj-GrootTeam_Groot-19B-A3B-261023-q8_0.gguf \
  --jinja -ngl 99 -c 32768

With MTP:

llama-server \
  -m Groot-19B-A3B-261023.APEX-I-MiniPlus-V2.1.gguf \
  --mmproj mmproj-GrootTeam_Groot-19B-A3B-261023-q8_0.gguf \
  --jinja -ngl 99 -c 32768 --spec-type draft-mtp

The Q8_0 projector is the smaller default; use the F16 file if you prefer the original projector precision.

The model starts in thinking mode. Disable it per request with the chat template kwargs, for example enable_thinking: false, if your client exposes them.

Recommended sampling

Use min_p = 0.0 and repetition_penalty = 1.0.

Limitations

• The base model benchmarks are self-reported by GrootTeam. They were not independently reproduced for this quantized release.
• Vision was not re-evaluated after quantization. The Q8_0 projector was generated from the original F16 projector, but vision quality was not tested.
• MTP speculative decoding was not benchmarked.
• GrootTeam does not disclose the exact method used to scale the 35B base down to 19B. Evaluate the model on your workload.
• Groot requires llama.cpp b12017 or newer.
• This release is not affiliated with GrootTeam.

License

Apache-2.0, inherited from the base model. See the base model card for the original license and terms.