NeoHorse-1-9B GGUF

Community GGUF quantizations of TokenRhythm/NeoHorse-1-9B.

☕ If this GGUF made your day easier, a coffee would make mine.
Send a coffee ☕
I build and test these releases myself. Your coffee helps keep me going.
Thank you for supporting this work.

About NeoHorse-1-9B

NeoHorse-1-9B is an approximately 9B causal language model post-trained from Qwen3.5-9B for text-based agent harnesses, tool use, coding, reasoning, and instruction following. This package contains text-only GGUF files; vision weights are not included. See the official upstream model card for the model's original details, usage, and limitations.

The upstream card describes a native 262,144-token context and extensibility up to 1,010,000 tokens. The upstream card also reports a 69.04 ten-benchmark average versus 65.60 for Qwen3.5-9B under its SGLang evaluation protocol. Those figures are upstream measurements, not measurements of these GGUF files.

NeoHorse-1-9B evaluation results (upstream)

This is a quantization-only release. No training or fine-tuning was performed as part of this conversion.

Fidelity measurements

The table compares every published quantization with the converted BF16 reference on the same hold-out split. Evaluation used the first two chunks of wiki.valid.raw, 2,048 tokens per chunk, and the CUDA build of llama.cpp on an NVIDIA A10M. The BF16 baseline mean PPL was 5.641047. Values are averaged over the two evaluated chunks. File sizes use decimal GB (bytes / 1,000,000,000). Bold rows are practical sweet spots: Q4_K_M is the balanced default, Q5_K_M and Q5_K_S provide more fidelity headroom, and IQ4_XS is a compact alternative.

File	Size (GB)	Mean KLD ↓	Top-1 vs BF16 ↑	ΔPPL	RMS Δp
NeoHorse-1-9B-Q8_0.gguf	9.528	0.006655	98.436%	+0.732%	0.811%
NeoHorse-1-9B-Q6_K.gguf	7.359	0.009702	97.312%	+1.120%	1.408%
NeoHorse-1-9B-Q5_K_M.gguf	6.468	0.030926	94.673%	+2.158%	5.841%
NeoHorse-1-9B-Q4_K_M.gguf	5.629	0.029467	94.428%	+0.951%	4.429%
NeoHorse-1-9B-Q3_K_M.gguf	4.624	0.069233	88.368%	+4.573%	7.326%
NeoHorse-1-9B-Q2_K.gguf	3.827	0.199997	82.063%	+14.167%	12.510%
NeoHorse-1-9B-IQ2_XS.gguf	3.285	0.342380	77.761%	+30.698%	18.365%
NeoHorse-1-9B-IQ1_M.gguf	2.877	0.807561	64.565%	+96.977%	28.701%
NeoHorse-1-9B-Q1_0.gguf	1.966	15.570850	2.933%	+596289635.160%	63.365%
NeoHorse-1-9B-Q5_K_S.gguf	6.305	0.030149	95.112%	+1.381%	5.817%
NeoHorse-1-9B-Q4_K_S.gguf	5.352	0.030445	93.842%	+1.040%	4.639%
NeoHorse-1-9B-Q3_K_L.gguf	4.926	0.062186	89.541%	+3.793%	6.790%
NeoHorse-1-9B-Q3_K_S.gguf	4.259	0.131594	85.386%	+6.965%	9.989%
NeoHorse-1-9B-Q2_K_S.gguf	3.697	0.240966	81.183%	+21.317%	14.223%
NeoHorse-1-9B-IQ4_XS.gguf	5.196	0.026534	93.891%	+2.125%	3.807%
NeoHorse-1-9B-IQ4_NL.gguf	5.418	0.029899	93.695%	+3.189%	4.788%
NeoHorse-1-9B-IQ3_M.gguf	4.415	0.064802	88.905%	+2.843%	7.357%
NeoHorse-1-9B-IQ3_S.gguf	4.371	0.081996	89.541%	+2.072%	7.758%
`ΔPPL` is the percentage change in perplexity relative to BF16. Lower KLD,
ΔPPL, and RMS Δp, together with higher Top-1 agreement, indicate closer
next-token behavior to BF16. These are next-token fidelity measurements, not a
task benchmark; reasoning, coding, and tool-use behavior can vary by workload.
The machine-readable results are in `reproducibility/quality-summary.tsv`,
with the evaluation details in `reproducibility/manifest.md`.

Quick start

Q4_K_M is a practical starting point for local use. Q5_K_M, Q6_K, or Q8_0 provide more headroom for fidelity when memory allows. Q3_K_M and lower are low-memory experimental options; the measurements above show increasing distance from BF16.

./llama-cli \
  -m NeoHorse-1-9B-Q4_K_M.gguf \
  --chat-template-file chat_template.jinja \
  --jinja \
  --reasoning off \
  -p 'Answer briefly in English: What is GGUF and why is it useful for running language models locally?' \
  -n 128 -c 4096 -ngl 99

The included chat_template.jinja is retained from the upstream text-only checkpoint. Adjust context size, GPU layers, and batch settings for the available hardware.

Reproducibility and validation

Source: upstream revision ba5b6e4.
The BF16 GGUF was converted directly from the upstream BF16 safetensors with --no-nextn, because this text-only checkpoint has no MTP tensors even though its config retains an MTP layer-count field.
Every published file was quantized directly from that BF16 GGUF with the model-specific imatrix included under calibration/; no re-quantization from another quantized file was used.
All published files passed load and short English generation smoke tests; the BF16 source additionally passed --check-tensors. Runtime readings are summarized in reproducibility/runtime-summary.tsv.
Verify artifacts with SHA256SUMS.txt and reproduce the provenance from reproducibility/manifest.md.

Raw conversion, calibration, quantization, smoke-test, fidelity, and benchmark logs are kept local and are not part of this public model package.

License and attribution

The upstream model and these community quantizations are released under the Apache License 2.0, with the upstream LICENSE retained in this repository.