Recipe
gemma-4-26b-a4b-it-ud-q8-k-xl-intel-arc-pro-b70-32gb-llama-cpp-tp1
gemma-4-26b-a4b-it-ud-q8-k-xl-intel-arc-pro-b70-32gb-llama-cpp-tp1Observed LocalMaxxing leaderboard run. Evidence for compatibility, not an executable launch contract.
Record
- Status
- candidate
- Source
- localmaxxing
- Engine
- llama.cpp
- Engine version
- 9769 (c926ad098), IntelLLVM 2026.0.0
- Accelerators
- 1
- Tensor parallel
- 1
- Context tokens
- 32,768
- Max concurrency
- 1
- chat
- unknown
- reasoning
- unknown
- tools
- unknown
- vision
- unknown
Hugging Face model card
Identity
https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF- Repository
- unsloth/gemma-4-26B-A4B-it-GGUF
- Status
- known
- Link type
- Exact Hub repository
Public Hugging Face repository confirmed by the Hub API.
Candidate: useful compatibility or speed evidence. The registry does not offer Run until promotion requirements are met.
Observed configuration
llama.cpp
Evidence only · candidate · reference
Candidate evidence — not a Run contract
Observed source tokens
Mechanical split of the source command. Unverified against the engine CLI. The registry does not offer Run for this recipe.
scripts/serve-gemma4-26b-q8-production.sh;python3scripts/bench-openai-long-context-suite.py--base-urlhttp://127.0.0.1:19350--modelgemma4-26b-a4b-q8--suiterepro/gemma4-26b-a4b-q8-b70/long-context-suite-v1.json--case-idlc-24000-late--max-tokens96
| Flag | Value |
|---|---|
--base-url | http://127.0.0.1:19350 |
--model | gemma4-26b-a4b-q8 |
--suite | repro/gemma4-26b-a4b-q8-b70/long-context-suite-v1.json |
--case-id | lc-24000-late |
--max-tokens | 96 |
Environment
| Variable | Value |
|---|---|
GEMMA4_26B_PROFILE | service |
GPU_INDEX | 0 |
PORT | 19350 |
Measured speed
| Concurrency | Context | Prefill | Decode | TTFT ms | Status | Sweep |
|---|---|---|---|---|---|---|
| 1 | 32,768 | 996.6 | 115.2 | 32,682.1 | observed | gemma-4-26b-a4b-it-ud-q8-k-xl-intel-arc-pro-b70-32gb-llama-cpp-tp1-sweep |
Remaining fields
Identity, launch, related records, and measured speed are shown above. This is the rest of the normalized record.
- hardware count
- 1
- hardware id
- intel-arc-pro-b70-32gb
- id
- gemma-4-26b-a4b-it-ud-q8-k-xl-intel-arc-pro-b70-32gb-llama-cpp-tp1
- model instance id
- unsloth-gemma-4-26b-a4b-it-gguf--ud-q8-k-xl
- recipe source
- localmaxxing
- schema version
- local-ai-registry/v1
- speed sweep ids
- gemma-4-26b-a4b-it-ud-q8-k-xl-intel-arc-pro-b70-32gb-llama-cpp-tp1-sweep
- status
- candidate
capabilities
engine
- name
- llama.cpp
- version
- 9769 (c926ad098), IntelLLVM 2026.0.0
serving
- max concurrency
- 1
- max context tokens
- 32,768
- tensor parallel
- 1
Provenance & metadata (3)
facts
- capabilities.chat · provenance · captured at
- 2026-08-30T09:10:02Z
capabilities.chat · reason capability-not-verified
capabilities.chat · state unknown
- capabilities.reasoning · provenance · captured at
- 2026-08-30T09:10:02Z
capabilities.reasoning · reason capability-not-verified
capabilities.reasoning · state unknown
- capabilities.tools · provenance · captured at
- 2026-08-30T09:10:02Z
capabilities.tools · reason capability-not-verified
capabilities.tools · state unknown
- capabilities.vision · provenance · captured at
- 2026-08-30T09:10:02Z
capabilities.vision · reason capability-not-verified
capabilities.vision · state unknown
- engine.graph mode · provenance · captured at
- 2026-08-30T09:10:02Z
engine.graph mode · reason runtime-detail-not-published
engine.graph mode · state unknown
- serving.kv cache tokens · provenance · captured at
- 2026-08-30T09:10:02Z
serving.kv cache tokens · reason kv-cache-capacity-not-published
serving.kv cache tokens · state unknown
metadata
- localmaxxing · hardware label
- Intel Arc Pro B70
- localmaxxing · notes
- Gemma 4 26B A4B Q8 long-context service smoke on one Intel Arc Pro B70 using llama.cpp/SYCL. This is a prompt-processing and near-32K long-context decode result, not the short-context 124.977 tok/s LocalMaxxing record. Shape: fixed deterministic long-context suite case lc-24000-late, actual prompt_tokens=32571, completion_tokens=76, contextLength=32768, batchSize=1, temperature=0, seed=1, cached_tokens=0, exact JSON retrieval fields passed, unique prompt hash. tokSPrefill is prompt_tokens / TTFT approximation; tokSOut is generated-token decode throughput after TTFT; tokSTotal is (prompt_tokens + output_tokens) / wall time. MTP/speculation is draft-only from local Q4_0 MTP draft; accepted tokens are verified by the UD-Q8_K_XL target/verifier. No prompt cache, KV cache reuse, context checkpoints, response reuse, n-gram/history acceleration, or warmed repeated prompt is used. Supporting service ladder on 2026-07-02 passed 32/32 long-context rows across four independent one-B70 lanes with cached_tokens=0 and exact JSON validation; average lane median prefill was 1192.965 tok/s and average lane median long-context decode was 131.786 tok/s. Evidence: https://github.com/steveseguin/b70-optimization-lab/blob/8723f783fa904092dcc019272f7abff939879f39/data/gemma4-26b-prod-service-32k-smoke-20260703T002811Z.json and deployment notes at https://github.com/steveseguin/b70-optimization-lab/blob/8723f783fa904092dcc019272f7abff939879f39/results/gemma4-26b-a4b-q8-b70/production-service.md
- localmaxxing · observed command
- GPU_INDEX=0 PORT=19350 GEMMA4_26B_PROFILE=service scripts/serve-gemma4-26b-q8-production.sh; python3 scripts/bench-openai-long-context-suite.py --base-url http://127.0.0.1:19350 --model gemma4-26b-a4b-q8 --suite repro/gemma4-26b-a4b-q8-b70/long-context-suite-v1.json --case-id lc-24000-late --max-tokens 96
- localmaxxing · run id
- cmr47ivql0045nv011pfdjlaa
- localmaxxing · tokenized · arguments
- scripts/serve-gemma4-26b-q8-production.sh;, python3, scripts/bench-openai-long-context-suite.py, --base-url, http://127.0.0.1:19350, --model, gemma4-26b-a4b-q8, --suite, repro/gemma4-26b-a4b-q8-b70/long-context-suite-v1.json, --case-id, lc-24000-late, --max-tokens, 96
- localmaxxing · tokenized · environment · GEMMA4 26B PROFILE
- service
- localmaxxing · tokenized · environment · GPU INDEX
- 0
- localmaxxing · tokenized · environment · PORT
- 19350
- localmaxxing · tokenized · fidelity
- faithful
provenance
- captured at
- 2026-08-30T09:10:02Z
sources
| captured at | kind | url |
|---|---|---|
| 2026-08-30T09:10:02Z | normalized-recipe | www.localmaxxing.com/en/runs/cmr47ivql0045nv011pfdjlaa ↗ |