Recipe

gemma-4-31b-v2-int4-apple-m5-pro-48gb-mlx-tp1

gemma-4-31b-v2-int4-apple-m5-pro-48gb-mlx-tp1

Record

Status
candidate
Source
localmaxxing
Engine
mlx
Engine version
mlx-vlm 0.6.3 / mlx 0.32.0.dev20260702+e9463bb (qmv_wide)
Accelerators
1
Tensor parallel
1
chat
unknown
reasoning
unknown
tools
unknown
vision
unknown

Hugging Face model card

Identity

https://huggingface.co/jdfelo/gemma-4-31B-v2-MLX-4bit
Repository
jdfelo/gemma-4-31B-v2-MLX-4bit
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

mlx

Evidence only · candidate · reference

Candidate evidence — not a Run contract

Source
https://www.localmaxxing.com/en/runs/cmr5ezpgk009mq901bsesdvu6

No tokenized launch fields. This is measured or documented compatibility, not a Docker launch.

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
12,048250.733.4596historicalgemma-4-31b-v2-int4-apple-m5-pro-48gb-mlx-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
apple-m5-pro-48gb
id
gemma-4-31b-v2-int4-apple-m5-pro-48gb-mlx-tp1
model instance id
jdfelo-gemma-4-31b-v2-mlx-4bit--int4
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
gemma-4-31b-v2-int4-apple-m5-pro-48gb-mlx-tp1-sweep
status
candidate

capabilities

engine

name
mlx
version
mlx-vlm 0.6.3 / mlx 0.32.0.dev20260702+e9463bb (qmv_wide)

serving

tensor parallel
1
Provenance & metadata (3)

facts

capabilities.chat · provenance · captured at
2026-08-27T06:04:13.773Z
capabilities.chat · reason capability-not-verified
capabilities.chat · state unknown
capabilities.reasoning · provenance · captured at
2026-08-27T06:04:13.773Z
capabilities.reasoning · reason capability-not-verified
capabilities.reasoning · state unknown
capabilities.tools · provenance · captured at
2026-08-27T06:04:13.773Z
capabilities.tools · reason capability-not-verified
capabilities.tools · state unknown
capabilities.vision · provenance · captured at
2026-08-27T06:04:13.773Z
capabilities.vision · reason capability-not-verified
capabilities.vision · state unknown
description · provenance · captured at
2026-08-27T06:04:13.773Z
description · reason not-observed
description · state unknown
engine.graph mode · provenance · captured at
2026-08-27T06:04:13.773Z
engine.graph mode · reason runtime-detail-not-published
engine.graph mode · state unknown
metadata.localmaxxing.engine flags.attentionBackend · provenance · captured at
2026-08-27T06:04:13.773Z
metadata.localmaxxing.engine flags.attentionBackend · reason not-observed
metadata.localmaxxing.engine flags.attentionBackend · state unknown
metadata.localmaxxing.engine flags.flashAttn · provenance · captured at
2026-08-27T06:04:13.773Z
metadata.localmaxxing.engine flags.flashAttn · reason not-observed
metadata.localmaxxing.engine flags.flashAttn · state unknown
metadata.localmaxxing.engine flags.gpuLayers · provenance · captured at
2026-08-27T06:04:13.773Z
metadata.localmaxxing.engine flags.gpuLayers · reason not-observed
metadata.localmaxxing.engine flags.gpuLayers · state unknown
metadata.localmaxxing.engine flags.kvCacheDtype · provenance · captured at
2026-08-27T06:04:13.773Z
metadata.localmaxxing.engine flags.kvCacheDtype · reason not-observed
metadata.localmaxxing.engine flags.kvCacheDtype · state unknown
metadata.localmaxxing.engine flags.tensorParallel · provenance · captured at
2026-08-27T06:04:13.773Z
metadata.localmaxxing.engine flags.tensorParallel · reason not-observed
metadata.localmaxxing.engine flags.tensorParallel · state unknown
serving.kv cache tokens · provenance · captured at
2026-08-27T06:04:13.773Z
serving.kv cache tokens · reason kv-cache-capacity-not-published
serving.kv cache tokens · state unknown
serving.max concurrency · provenance · captured at
2026-08-27T06:04:13.773Z
serving.max concurrency · reason not-observed
serving.max concurrency · state unknown
serving.max context tokens · provenance · captured at
2026-08-27T06:04:13.773Z
serving.max context tokens · reason not-observed
serving.max context tokens · state unknown

metadata

localmaxxing · base lineage
derived
localmaxxing · engine flags · mtpEnabled
Yes
localmaxxing · engine flags · specDecoding
Yes
localmaxxing · hardware label
Apple M5 Pro 48GB
localmaxxing · notes
Gemma 4 31B (abliterated v2), 4-bit affine gs64 decoder; vision tower + embed/lm_head bf16 per config. MLX MTP speculative decoding via mlx-vlm, official gemma-4-31B-it-assistant drafter, draft-block-size 4 (optimum on this build). Measured on M5 Pro 48GB, macOS 26.5, greedy temp 0. tokSOut = median of 4 timed iterations (range 32.8-35.5), decode throughput over the inter-token window (completion_tokens / (t_last-t_first)) from the OpenAI-compatible stream; ~250-tok code workload, exact usage.completion_tokens. Drafter acceptance ~69.8%, ~3.09 accepted tok/round. Byte-identical to dense greedy (fast-verify patch adds the missing gemma4 speculative_argmax_from_hidden hook: one batched norm->lm_head->softcap over K+1 positions). Speedup vs dense 4-bit (~15 tok/s, ~94% of the 276 GB/s measured-bandwidth ceiling) ~2.2x. peakVramGb is process phys_footprint_peak and includes MLX's Metal buffer cache growing under a 36GB wired limit; true resident weights+drafter+KV ~= 18-20GB (steady footprint 26.5GB). wheel: mlx qmv_wide (PR #3764) built from git main; without it b3 tops out ~24.6 tok/s. Model repo public; base jdfelo/gemma-4-31B-v2 is gated.
localmaxxing · revision unpinned
Yes
localmaxxing · run id
cmr5ezpgk009mq901bsesdvu6

provenance

captured at
2026-08-27T06:04:13.773Z

sources

captured atkindurl
2026-08-27T06:04:13.773Znormalized-recipewww.localmaxxing.com/en/runs/cmr5ezpgk009mq901bsesdvu6