Spark-X2.5-4B — Pollard

Pollard shrank this model: 8.22 GB (f16) → 1.93 GB — 77% smaller, 4.3× down.

The smallest rung here; larger, higher-fidelity rungs are listed below.

format	this model's size
f16	8.22 GB
Q8_0	~4.36 GB
Q6_K	~3.37 GB
Q4_K_M	~2.38 GB
PollardMix (this repo's IQ3_S)	1.93 GB

Pollard builds of XHToken/Spark-X2.5-4B made with Pollard Weights — a ladder of measured-allocation quants (bits placed by per-layer sensitivity, not a uniform crush).

Standard GGUF, but you need a recent llama.cpp. This model's architecture (spark2_5) is implemented upstream, so any build new enough to carry it runs these files -- llama.cpp itself, and Ollama or LM Studio once they ship a runtime with it. An older build will refuse them with unknown model architecture. The quants are ordinary K-quants.

Model details

Which file should I choose?

Every rung is the same weights, sized to a different RAM budget by the measured allocation. Pick the largest one that fits your machine with room for context:

• ~5 GB RAM / VRAM → Q6_K (3.38 GB). (needs a current llama.cpp) near-lossless
• ~4 GB RAM / VRAM → IQ4_XS (2.42 GB). (needs a current llama.cpp) recommended default
• ~4 GB RAM / VRAM → IQ3_S (1.93 GB). (needs a current llama.cpp) smallest

Available files (Calib 3.0 corpus, ctx 512, 6 chunks)

f16 reference PPL 6.091...

Measured notes

f16 reference PPL 6.091.

Measured notes

f16 reference PPL 6.091..

Download a specific file

pip install -U "huggingface_hub[cli]"
hf download PollardWeights/Spark-X2.5-4B-Pollard \
  --include "Spark-X2.5-4B-Pollard-IQ4_XS.gguf" --local-dir ./

How to run

This model's architecture (spark2_5) needs a llama.cpp new enough to carry it, so build or update from upstream first:

llama-cli    -m Spark-X2.5-4B-Pollard-IQ4_XS.gguf -ngl 99 -p "Explain why the sky is blue."
llama-server -m Spark-X2.5-4B-Pollard-IQ4_XS.gguf -ngl 99

Ollama and LM Studio will run these once their bundled llama.cpp carries spark2_5.

ARM / AVX

llama.cpp repacks weights into an interleaved layout at load time for faster inference on ARM and AVX machines — no special file needed, online repacking covers these quants. The old Q4_0_4_4/4_8/8_8 variants are not required.

Errata

• general.architecture is spark2_5, which upstream llama.cpp added recently. A build older than that support refuses these files with unknown model architecture — update llama.cpp rather than looking for a different quant. Checked with pollard-ggufcheck, which reads the architecture and the tensor types out of the header and asks upstream what it implements.
• Measured allocation places bits by per-layer sensitivity under a size budget.
• Single machine; replication invited.

Credits & license

• Base model: XHToken/Spark-X2.5-4B
• Quantization tooling: llama.cpp (ggml-org)
• Method + tooling: Pollard Weights — measure first, no claim before a number.
• License: apache-2.0, inherited from the base model.

Built with Pollard Weights — frontier models, small hardware, no compromise.