Modelpublic

VeigaPunk/gemma4-hvm-official-q4

sourceHugging Facegemmaupdated 2mo agoView on Hugging Face

0likes

Model Card

Gemma 4 26B A4B IT QAT Q4_0 — HVM/Bend lane

This repository packages Google's official Gemma 4 26B A4B instruction-tuned QAT Q4_0 GGUF used by the local gemma-hvm lane.

Important: HVM/Bend is the orchestration and native-call path, not the tensor inference engine. Model inference is performed locally by Ollama using this GGUF.

Model details

Property	Value
Architecture	Gemma 4 MoE
Parameters reported by Ollama	25.2B
Quantization	Q4_0
GGUF size	14,439,363,584 bytes (14.44 GB / 13.45 GiB)
SHA-256	`3eca3b8f6d7baf218a7dd6bba5fb59a56ee25fe2d567b6f5f589b4f697eca51d`
Upstream revision	`d1c082be9cf3c8a514acf63b8761f4b41935842e`
Ollama tag used locally	`gemma4-hvm:official-q4`

Local size comparison

Model	Parameters	Quantization	Ollama size
`gemma4:12b`	11.9B	Q4KM	7.6 GB
`gemma4:e4b`	8.0B	Q4KM	9.6 GB
`gemma4-hvm:official-q4`	25.2B	Q4_0	14 GB
`gemma4:26b`	25.8B	Q4KM	17 GB

The official Q4 model retains approximately the full 26B-class parameter count while using about 3 GB less Ollama storage than the local Q4KM gemma4:26b build.

HVM route

text

CLI / xbreed
  → gemma-hvm
  → Bend
  → HVM2
  → native DyLib bridge
  → Ollama /api/generate
  → local GGUF

The bridge supports environment-controlled model identity, context, output budget, temperature, seed, thinking mode, keep-alive, endpoint, timeout, and system prompt.

Download and verify

bash

hf download VeigaPunk/gemma4-hvm-official-q4 \
  gemma-4-26B_q4_0-it.gguf \
  --local-dir .

sha256sum gemma-4-26B_q4_0-it.gguf

Expected digest:

text

3eca3b8f6d7baf218a7dd6bba5fb59a56ee25fe2d567b6f5f589b4f697eca51d

Run with Ollama

The repository includes this Modelfile:

dockerfile

FROM ./gemma-4-26B_q4_0-it.gguf

TEMPLATE {{ .Prompt }}
PARAMETER num_ctx 2048
PARAMETER temperature 0
PARAMETER stop <turn|>

Create and run the model:

bash

ollama create gemma4-hvm:official-q4 -f Modelfile
ollama run gemma4-hvm:official-q4

Run through HVM/Bend

With the companion hvm-gemma4 project installed:

bash

HVM_GEMMA_MODEL=gemma4-hvm:official-q4 \
  gemma-hvm "Return only the exact string: pong"

Expected first output line:

text

pong

Tuned local defaults

The bounded local tuning campaign selected these defaults for an RTX 5070 12 GB / 30 GB RAM workstation:

Setting	Value
Model	`gemma4-hvm:official-q4`
Context	`2048`
General output budget	`256`
Concise benchmark profile	`64`
Temperature	`0`
Seed	`42`
Thinking	`false`
Keep-alive	`10m`
KV cache	`q8_0`
Ollama parallel slots	`8`

A concise format-following system instruction improved the deterministic calibration suite from 4/12 to 12/12 in local development. This is a small configuration-tuning benchmark, not a general capability score. The held-out validator set was inspected during development and must not be treated as an uncontaminated final evaluation.

Environment variables

text

HVM_GEMMA_MODEL
HVM_GEMMA_SYSTEM
HVM_GEMMA_NUM_CTX
HVM_GEMMA_NUM_PREDICT
HVM_GEMMA_TEMPERATURE
HVM_GEMMA_SEED
HVM_GEMMA_THINK
HVM_GEMMA_KEEP_ALIVE
HVM_GEMMA_HTTP_TIMEOUT
HVM_GEMMA_ENDPOINT

Limitations

—HVM does not execute the model tensors; it orchestrates the Bend/HVM/native bridge.
—The Q40 artifact may differ in quality, speed, and memory use from Q4K_M variants.
—The model metadata advertises a much larger context window, but the tested HVM profile uses num_ctx=2048 for predictable local resource use.
—Tool calling and thinking capabilities reported by the model metadata are not equivalent to full agent-tool execution through the plain-text HVM lane.
—Performance and memory behavior depend heavily on hardware, Ollama version, cache precision, concurrency, and model residency.

License and attribution

The model is derived from Google's Gemma release. Use, modification, and redistribution are subject to the Gemma Terms of Use and the upstream model license. Review and accept those terms before downloading or redistributing the weights.

Upstream model:

—https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-gguf