Model Follower.

Up to date

Rebuilt
2026-10-25 06:54 UTC
Release radar
117 on watch
Hub snapshot
2026-10-25
Re-checked
daily · every number sourced and dated

Models › gemma-4 family

gemma-4-12B-it

Hugging Face ↗ Follow changes

google Apache 2.0 out 5 months ago

12.0B parameters, several kinds of input and output, 256K tokens of context.

Verified 2026-10-25 Rank #106 in LLMs · score 98.7 method v1 Sources 3 Downloads 30d 1.7M · #73 of 619 tracked

Changed this week: 17 new quantized builds on the Hub (375 in all).2026-10-24 · 9 new quantized builds on the Hub (367 in all).2026-10-24

Can I run it?7.2GB at INT4Estimatefits a 16 GB cardmethodCommunity: gemma-4-12B-it: architecture, parameters and context length | Frontier ML →
Can I use it?Apache 2.0Permissive; patent grantlicence file2026-10-24
Do people use it?1.7Mdownloads, 30 days#73 of 619 we track · top 12%+14 likes, downloads +2% this week1,685 likesHugging Face2026-10-24

Worth reading

Picked by rule from 34 tracked pieces: the lab's own post, the most shared, a field report, the latest.

The lab's own post

Gemma 4 12B: A unified, encoder-free multimodal model

Google releases Gemma 4 12B, an Apache 2.0 licensed model that has crossed 150 million downloads and runs locally on hardware with 16GB of VRAM or unified memory.

blog.google 2026-06-03 1,062 HN pts · 401 comments 584 words
Most discussed · 868 HN pts

Gemma 4 on iPhone

apps.apple.com 2026-04-05 868 HN pts · 234 comments about the gemma 4 family family
Latest · 3 months ago

Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong

Cactus Hybrid introduces on-device models that score every answer with a confidence between 0 and 1 to decide when to hand off to a bigger model.

  • Weights on the meta device crash the confidence read because probes score generations outside the module forward path.
  • The probe is a patch you compile into the llama.cpp engine rather than a standalone library component.
github.com 2026-07-22 191 HN pts · 44 comments 714 words about the gemma 4 family family

25 best of 34 tracked

The 3 picks above, plus these 22. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. CodeShow HN: Open-source engine running Gemma 4 26B in 2 GB RAM on any M-series Macgithub.com · 2026-07-29
  2. ArticleRunning Gemma 4 26B at 5 tokens/sec on a 13-year-old Xeon with no GPUneomindlabs.com · 2026-07-15
  3. ArticleGemma 4 QAT models: Optimizing compression for mobile and laptop efficiencyblog.google · 2026-06-05 · 406 HN pts · official
  4. ArticleAccelerating Gemma 4: faster inference with multi-token prediction draftersblog.google · 2026-05-05 · 687 HN pts · official
  5. ArticleGoogle releases Gemma 4 open modelsdeepmind.google · 2026-04-02 · 1,812 HN pts · official
  6. CodeGemma 4 E2B inference in 700 lines of Cgithub.com · 2026-08-27 · 18 HN pts
  7. ArticleFable 5 pushed Gemma 4 to 255 tok/s on WebGPUxcancel.com · 2026-06-18 · 48 HN pts
  8. ArticleGoogle Gemma 4 Runs Natively on iPhone with Full Offline AI Inferencegizmoweek.com · 2026-04-15 · 303 HN pts
  9. CodeApril 2026 TLDR Setup for Ollama and Gemma 4 26B on a Mac minigist.github.com · 2026-04-03
  10. ArticleShow HN: Determinstic LLM inference for lowest price Gemma 4, with Windows XPtokendelivery.ai · 2026-09-12 · 9 HN pts
  11. ArticleI ran Gemma 4 as a local model in Codex CLIblog.danielvaughan.com · 2026-04-12
  12. ArticleRunning Gemma 4 locally with LM Studio's new headless CLI and Claude Codeai.georgeliu.com · 2026-04-05
  13. ArticleGemma 4 on Cerebras - The Fastest Inference Is Now Multimodalcerebras.ai · 2026-06-30
  14. CodeShow HN: Gemma 4 Multimodal Fine-Tuner for Apple Silicongithub.com · 2026-04-07
  15. ArticleShow HN: Prompt-to-Excalidraw demo with Gemma 4 E2B in the browser (3.1GB)teamchong.github.io · 2026-04-19
  16. CodeShow HN: Local privacy-first Microsoft Recall alternative with Gemma 4github.com · 2026-07-04
  17. ArticleGemma 4: Byte for byte, the most capable open modelsblog.google · 2026-04-02 · official
  18. ArticleA Visual Guide to Gemma 4 12Bnewsletter.maartengrootendorst.com · 2026-06-03
  19. CodeShow HN: Local LLM code-generation with Gemma 4 e2B via JSON AST to Clojuregithub.com · 2026-05-19
  20. ArticleShow HN: Running Gemma-4 26B at 124 tokens/SEC on a CPU, no GPUapeg.dev · 2026-06-30
  21. ArticleShow HN: I made a Gemma 4 Mac app that names screenshots with local AIsnapname.app · 2026-05-31
  22. ArticleGemma 4 12B: The Developer Guidedevelopers.googleblog.com · undated · official

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

No API provider serves it, so its downloads are people running it themselves: 375 quantized builds and 90 Spaces.

  • As a service0no API provider listed
  • On own hardware375quantized builds
  • Built on404254 fine-tunes · 139 adapters · 11 merges
  • In apps90Hugging Face Spaces

Top Spaces running it: huggingface-projects/gemma-4-12b-it ♥ 156 · build-small-hackathon/claim-ready ♥ 71 · build-small-hackathon/tutori ♥ 19 · build-small-hackathon/jackailocal ♥ 6

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits a 16 GB card at INT4

About 7.2 GB at INT4, 14 GB at FP8, 29 GB at BF16: weights only, so leave room for the KV cache.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 7.2 GB of weights (estimate; KV cache and activations extra)INT47.2 GBFP8: about 14 GB of weights (estimate; KV cache and activations extra)FP814 GBBF16: about 29 GB of weights (estimate; KV cache and activations extra)BF1629 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 7.2 GB of weights (estimate; KV cache and activations extra)INT47.2 GBFP8: about 14 GB of weights (estimate; KV cache and activations extra)FP814 GBBF16: about 29 GB of weights (estimate; KV cache and activations extra)BF1629 GB

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

Most coverage came in the first two weeks

Teal is the weights (2026-05-23), amber triangles are 10 dated articles (each links to it), grey ticks are repo commits.

MayJunJulAugSepOct2026-07-20: 1 repo commit2026-07-15: 2 repo commits2026-07-10: 1 repo commit2026-06-04: 1 repo commit2026-06-03: 2 repo commitsGemma 4 12B: A unified, encoder-free multimodal model · 2026-06-03Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong · 2026-07-22Gemma 4 QAT models: Optimizing compression for mobile and laptop efficiency · 2026-06-05Accelerating Gemma 4: faster inference with multi-token prediction drafters · 2026-05-05Google releases Gemma 4 open models · 2026-04-02Gemma 4 E2B inference in 700 lines of C · 2026-08-27Fable 5 pushed Gemma 4 to 255 tok/s on WebGPU · 2026-06-18Gemma 4 on iPhone · 2026-04-05Google Gemma 4 Runs Natively on iPhone with Full Offline AI Inference · 2026-04-15Show HN: Determinstic LLM inference for lowest price Gemma 4, with Windows XP · 2026-09-12Weights 2026-05-23 · 2026-05-23Weights 2026-05-23now
MayJunJulAugSepOct2026-07-20: 1 repo commit2026-07-15: 2 repo commits2026-07-10: 1 repo commit2026-06-04: 1 repo commit2026-06-03: 2 repo commitsGemma 4 12B: A unified, encoder-free multimodal model · 2026-06-03Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong · 2026-07-22Gemma 4 QAT models: Optimizing compression for mobile and laptop efficiency · 2026-06-05Accelerating Gemma 4: faster inference with multi-token prediction drafters · 2026-05-05Google releases Gemma 4 open models · 2026-04-02Gemma 4 E2B inference in 700 lines of C · 2026-08-27Fable 5 pushed Gemma 4 to 255 tok/s on WebGPU · 2026-06-18Gemma 4 on iPhone · 2026-04-05Google Gemma 4 Runs Natively on iPhone with Full Offline AI Inference · 2026-04-15Show HN: Determinstic LLM inference for lowest price Gemma 4, with Windows XP · 2026-09-12Weights 2026-05-23 · 2026-05-23Weights 2026-05-23now

What people built on it

779 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

779 derivatives on the Hub, led by 375 quantized

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

375 quantized (Hugging Face model tree)254 fine-tunes (Hugging Face model tree)11 merges (Hugging Face model tree)139 adapters (Hugging Face model tree)375 quantized254 fine-tunes11 merges139 adapters
375 quantized (Hugging Face model tree)254 fine-tunes (Hugging Face model tree)11 merges (Hugging Face model tree)139 adapters (Hugging Face model tree)375 quantized254 fine-tunes11 merges139 adapters
GGUF unsloth/gemma-4-12b-it-GGUF for llama.cpp, Ollama, LM Studio 5.3M downloads MLX lmstudio-community/gemma-4-12B-it-MLX-4bit for Apple silicon 1.1M downloads AWQ cyankiwi/gemma-4-12B-it-AWQ-INT4 for vLLM, SGLang 1.5M downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

Benchmarks

16 rows

Benchmarks 16 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (16)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.
BenchmarkValueSetting / notesReported byDate
author-reported AIME 2026 · Reason through hard problems77.5% percentGemma 4 12B Unifiedmodel author2026-10-07
author-reported BBEH · Reason through hard problems53.0% percentGemma 4 12B Unifiedmodel author2026-10-07
author-reported CoVoST38.5* otherGemma 4 12B Unifiedmodel author2026-10-07
author-reported Codeforces · Write code with agents1659 eloGemma 4 12B Unifiedmodel author2026-10-07
author-reported FLEURS (lower is better)0.069* otherGemma 4 12B Unifiedmodel author2026-10-07
author-reported GPQA Diamond · Reason through hard problems78.8% percentGemma 4 12B Unifiedmodel author2026-10-07
author-reported Humanity's Last Exam · Reason through hard problems5.2% percentGemma 4 12B Unifiedmodel author2026-10-07
author-reported LiveCodeBench · Write code with agents72.0% percentGemma 4 12B Unifiedmodel author2026-10-07
author-reported MMLU-Pro · Reason through hard problems77.2% percentGemma 4 12B Unifiedmodel author2026-10-07
author-reported MMMLU · Work across languages83.4% percentGemma 4 12B Unifiedmodel author2026-10-07
author-reported MMMU-Pro · Understand images and video69.1% percentGemma 4 12B Unifiedmodel author2026-10-07
author-reported MRCR · Read long documents43.4% percentGemma 4 12B Unifiedmodel author2026-10-07
author-reported MathVision · Understand images and video79.7% percentGemma 4 12B Unifiedmodel author2026-10-07
author-reported MedXPertQA MM48.7% percentGemma 4 12B Unifiedmodel author2026-10-07
author-reported OmniDocBench · Extract text from documents0.16 otherGemma 4 12B Unifiedmodel author2026-10-07
author-reported tau-bench · Call tools and APIs69.0% percentGemma 4 12B Unifiedmodel author2026-10-07

What the card says

3 cited sentences

Summary generated, cited

Gemma 4 12B-it is an instruction-tuned, 12.0B parameter multimodal model from Google DeepMind.[1][2] The model uses a Gemma4UnifiedForConditionalGeneration architecture to process text, image, and audio inputs.[3][4] It is released under the Apache 2.0 license and supports a context window of up to 256K tokens.[5][6]

  1. parameters human
  2. “instruction-tuned”
  3. architectures
  4. “Text, Image, Audio”
  5. license
  6. “context window of up to 256K tokens”

Generated 2026-10-07 15:01 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

9 use cases

What it is good at 9 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

21 fields

Listed in LLMs

Facts

Pipeline tagany-to-any
Librarytransformers
ArchitecturesGemma4UnifiedForConditionalGeneration
Model typegemma4_unified
Dtypebfloat16
Params by dtypeBF16: 11,959,730,224
Size bandL
MoEno
Context length262,144 (config.json max_position_embeddings)
PapersarXiv:2607.02770
LicenseApache 2.0 (apache-2.0)
Hub snapshot2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · Gemma4UnifiedForConditionalGeneration · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · Gemma4UnifiedForConditionalGeneration · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for gemma-4-12B-it: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)29 GB · Single high-end GPU (<80 GB)
FP8 (1 byte/param)14 GB · Single GPU (<24 GB)
INT4 (0.5 bytes/param)7.2 GB · Single GPU (<24 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.

All variants

40 variants

Lineage

Fine-tuned from google/gemma-4-12B (same organization) Family: google/gemma-4 779 derivatives on the Hub

Base model: google/gemma-4-12B

quantized: 375finetune: 254merge: 11adapter: 139

Top variants (40 shown, ranked by downloads)

Relation
Quant
Size
40 rows
Variant Relation Quant Params Band License Author Likes Downloads
unsloth/gemma-4-12b-it-GGUF quantized GGUF — — apache-2.0 unsloth 929 5.3M
yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF quantized GGUF — — apache-2.0 yuxinlu1 1.7k 2.2M
yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF quantized GGUF — — apache-2.0 yuxinlu1 3k 2M
HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced quantized GGUF — — gemma HauhauCS 471 2.1M
cyankiwi/gemma-4-12B-it-AWQ-INT4 quantized AWQ 12.0B L apache-2.0 cyankiwi 9 1.5M
lmstudio-community/gemma-4-12B-it-GGUF quantized GGUF — — apache-2.0 lmstudio-community 34 1.1M
lmstudio-community/gemma-4-12B-it-MLX-4bit quantized MLX 12.0B L apache-2.0 lmstudio-community 7 1.1M
google/gemma-4-12B-it-qat-q4_0-unquantized finetune q4_0 12.0B L apache-2.0 google 97 1.1M
lmstudio-community/gemma-4-12B-it-MLX-8bit quantized MLX 12.0B L apache-2.0 lmstudio-community 6 1.1M
Comfy-Org/gemma-4 finetune — — — apache-2.0 Comfy-Org 109 999.1k
lmstudio-community/gemma-4-12B-it-MLX-6bit quantized MLX 12.0B L apache-2.0 lmstudio-community 2 1M
lmstudio-community/gemma-4-12B-it-MLX-5bit quantized MLX 12.0B L apache-2.0 lmstudio-community 0 988.7k
RedHatAI/gemma-4-12B-it-NVFP4 quantized NVFP4 13.0B L apache-2.0 RedHatAI 6 899.9k
RedHatAI/gemma-4-12B-it-FP8-Dynamic quantized -FP8 13.0B L apache-2.0 RedHatAI 9 696.3k
ggml-org/gemma-4-12B-it-GGUF quantized GGUF — — apache-2.0 ggml-org 94 583k
rahul7star/gemma-gguf quantized gguf — — — rahul7star 2 526.8k
unsloth/gemma-4-12b-it finetune — 12.0B L apache-2.0 unsloth 20 412.9k
yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2 finetune — 12.0B L apache-2.0 yuxinlu1 109 337.4k
OBLITERATUS/Gemma-4-12B-OBLITERATED quantized GGUF 12.0B L gemma OBLITERATUS 478 317.4k
bartowski/gemma-4-12B-it-GGUF quantized GGUF — — apache-2.0 bartowski 52 315k
mattbucci/gemma-4-12B-AWQ quantized AWQ 12.0B L gemma mattbucci 4 220.6k
xCloudinfo/xGemable-12B-coder-v1.5-GGUF quantized GGUF — — apache-2.0 xCloudinfo 1 120.8k
unsloth/gemma-4-12b-it-NVFP4 quantized NVFP4 8.24B M apache-2.0 unsloth 26 106.9k
coolthor/gemma-4-12B-it-FP8-dynamic quantized -FP8 12.0B L apache-2.0 coolthor 2 98.7k
yuxinlu1/gemma-4-12B-it-Claude-4.6-4.8-Opus-GGUF quantized GGUF — — apache-2.0 yuxinlu1 131 90.9k
litert-community/gemma-4-12B-it-litert-lm finetune — — — apache-2.0 litert-community 67 85.7k
toandev/Gemma4-12B-Uncensored finetune — 12.0B L gemma toandev 3 84.5k
AxionML/Gemma-4-12B-NVFP4 quantized NVFP4 7.71B M apache-2.0 AxionML 8 66.6k
AxionML/Gemma-4-12B-FP8 quantized -FP8 12.0B L apache-2.0 AxionML 6 62.1k
huihui-ai/Huihui-gemma-4-12B-it-abliterated finetune — 12.0B L apache-2.0 huihui-ai 226 37.5k
mlx-community/gemma-4-12B-it-OptiQ-4bit quantized -4bit 11.9B L gemma mlx-community 56 43.9k
yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1 finetune — 12.0B L apache-2.0 yuxinlu1 65 41.3k
OpenYourMind/gemma-4-12B-it-abliterated-uncensored finetune — 12.0B L gemma OpenYourMind 81 21.8k
unsloth/gemma-4-12B-it-qat-q4_0-unquantized finetune q4_0 12.0B L apache-2.0 unsloth 9 22k
SC117/Gemma-4-12B-it-heretic-GGUF quantized GGUF — — apache-2.0 SC117 3 20k
AEON-7/Gemma-4-12B-it-AEON-Abliterated-K4-BF16 finetune -BF16 12.0B L gemma AEON-7 32 18.1k
w-ahmad/Gemma-4-12B-it-qat-GGUF-MoQ quantized GGUF — — apache-2.0 w-ahmad 13 18.3k
prithivMLmods/ultragemma4-12b-heretic-uncensored quantized GGUF — — apache-2.0 prithivMLmods 4 18.1k
Intel/gemma-4-12B-it-int4-AutoRound quantized int4 2.52B M — Intel 4 15.7k
FreedomAISVR/Gemma-4-12B-it-NVFP4-GGUF quantized NVFP4 — — apache-2.0 FreedomAISVR 6 14.6k

Repository changelog

7 commits
changed in the last 7 daysRank band moves: 2026-10-24, 2026-10-23Other: 2026-10-24, 2026-10-24all changes

Repository changelog 7 commits

DateCommitSHA
2026-07-20 Add response_template to tokenizer_config.json (#43) 707f0a3
2026-07-15 Update README (#47) 12ace6d
2026-07-15 fix: chat template — null handling, reasoning preservation, turn-tag balance, input validation (#35) 711c136
2026-07-10 Add Technical Report 0e2b105
2026-06-04 Update max_position_embeddings to full context length (#21) 5926caa
2026-06-03 Update README.md 66bc78a
2026-06-03 Super-squash branch 'main' using huggingface_hub 657684f

Articles and coverage

34 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official4

Gemma 4 12B: A unified, encoder-free multimodal model
blog.google2026-06-03officialauthorityHN front pageHN 1,062 pts · 401 comments584 words
Google releases Gemma 4 12B, an Apache 2.0 licensed model that has crossed 150 million downloads and runs locally on hardware with 16GB of VRAM or unified memory. The encoder-free architecture integrates raw audio signals and vision inputs directly into the LLM backbone, eliminating separate encoders to reduce latency and memory usage. It includes Multi-Token Prediction (MTP) drafters to reduce inference latency while delivering reasoning performance nearing the 26B Mixture of Experts model.written, reviewed, cited
Gemma 4 QAT models: Optimizing compression for mobile and laptop efficiency
blog.google2026-06-05officialauthorityHN front pageabout the gemma 4 familyHN 406 pts · 130 comments714 words
Google released new Gemma 4 checkpoints optimized with Quantization-Aware Training to improve efficiency on edge devices and consumer GPUs. The E2B model achieves a 1GB memory footprint using a mobile-specialized schema that applies targeted 2-bit quantization to token generation layers. Developers can access GGUF formats ready for llama.cpp and MTP QAT checkpoints to preserve inference speedups while running locally.written, reviewed, cited
  • The mobile-specialized schema pre-calculates activation scaling settings during training to reduce workload on mobile chips.
  • Targeted 2-bit quantization compresses token generation parts while keeping core reasoning layers at higher precision.
  • Deploying only needed modalities allows the Gemma 4 E2B text-only model to require less than 1 GB of memory.
Accelerating Gemma 4: faster inference with multi-token prediction drafters
blog.google2026-05-05officialauthorityHN front pageabout the gemma 4 familyHN 687 pts · 330 comments796 words
Google released Multi-Token Prediction (MTP) drafters for the Gemma 4 family under the same open-source Apache 2.0 license as Gemma 4. These drafters deliver up to a 3x speedup for models like the 26B MoE and 31B Dense versions without quality degradation. Benchmarks on hardware such as the NVIDIA RTX PRO 6000 show that increasing batch sizes to 4 to 8 unlocks up to a ~2.2x speedup locally.written, reviewed, cited
  • The MTP drafter shares the target model's KV cache to avoid recalculating context the larger model has already processed.
  • For E2B and E4B edge models, an efficient clustering technique in the embedder accelerates generation where logit calculation is a bottleneck.
  • The 26B mixture-of-experts model presents unique routing challenges at a batch size of 1 on Apple Silicon.
Google releases Gemma 4 open models
deepmind.google2026-04-02officialauthorityHN front pageabout the gemma 4 familyHN 1,812 pts · 474 comments253 words

Analysis6

Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong
github.com2026-07-22authorityHN front pageabout the gemma 4 familyHN 191 pts · 44 comments714 words
Cactus Hybrid introduces on-device models that score every answer with a confidence between 0 and 1 to decide when to hand off to a bigger model. The Gemma 4 E2B Hybrid model matches Gemini 3.1 Flash-Lite performance by routing only 15–55% of queries to the cloud while running the rest locally. It is MIT-licensed and achieves a 0.770 AUROC on MMLU, with a probe trained on zero audio data yet reaching 0.79–0.88 AUROC on four audio benchmarks.written, reviewed, cited
  • Weights on the meta device crash the confidence read because probes score generations outside the module forward path.
  • The probe is a patch you compile into the llama.cpp engine rather than a standalone library component.
  • Transformers version 5.14 and above cause a segfault on this specific checkpoint.
Gemma 4 E2B inference in 700 lines of C
github.com2026-08-27authorityabout the gemma 4 familyHN 18 pts · 0 comments722 words
Fable 5 pushed Gemma 4 to 255 tok/s on WebGPU
xcancel.com2026-06-18about the gemma 4 familyHN 48 pts · 22 comments
Gemma 4 on iPhone
apps.apple.com2026-04-05HN front pageabout the gemma 4 familyHN 868 pts · 234 comments
Google Gemma 4 Runs Natively on iPhone with Full Offline AI Inference
gizmoweek.com2026-04-15HN front pageabout the gemma 4 familyHN 303 pts · 187 comments364 words
Show HN: Determinstic LLM inference for lowest price Gemma 4, with Windows XP
tokendelivery.ai2026-09-12about the gemma 4 familyHN 9 pts · 1 comments
All coverage tracked for this model (34) — append-only, nothing is ever removed
TitleSourcePublishedFirst seenSignals
Gemma 4 12B: The Developer Guidedevelopers.googleblog.com official—2026-10-07exa
Gemma 4 Technical Report - arXivarxiv.org—2026-10-07exa
gemma-4-12B-it: architecture, parameters and context length | Frontier MLfrontiermlnp.com—2026-10-07exa
Show HN: Determinstic LLM inference for lowest price Gemma 4, with Windows XPtokendelivery.ai2026-09-122026-10-07hn
Gemma 4 E2B inference in 700 lines of Cgithub.com2026-08-272026-10-07hn
Show HN: Open-source engine running Gemma 4 26B in 2 GB RAM on any M-series Macgithub.com2026-07-292026-10-07hn
Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wronggithub.com2026-07-222026-10-07hn
Running Gemma 4 26B at 5 tokens/sec on a 13-year-old Xeon with no GPUneomindlabs.com2026-07-152026-10-07hn
Show HN: Local privacy-first Microsoft Recall alternative with Gemma 4github.com2026-07-042026-10-07hn
Show HN: Running Gemma-4 26B at 124 tokens/SEC on a CPU, no GPUapeg.dev2026-06-302026-10-07hn
Gemma 4 on Cerebras - The Fastest Inference Is Now Multimodalcerebras.ai2026-06-302026-10-07hn
Fable 5 pushed Gemma 4 to 255 tok/s on WebGPUxcancel.com2026-06-182026-10-07hn
Gemma 4 QAT models: Optimizing compression for mobile and laptop efficiencyblog.google official2026-06-052026-10-07hn
Gemma 4 12B: 7 Proven Fixes For Setup, GGUF And Benchmarksbinaryverseai.com2026-06-042026-10-07exa
Gemma 4 12B Deep Dive — The Encoder-Free Multimodal LLM That Runs on a 16GB Laptop Under Apache 2.0 (June 3, 2026) | Oflight Inc.oflight.co.jp2026-06-042026-10-07exa
Gemma 4 12B : Run Locally, Fine-Tune, Benchmark Performancelabellerr.com2026-06-042026-10-07exa
Google's new Gemma 4 12B model is designed to run on any laptop with 16GB of RAMarstechnica.com2026-06-032026-10-07hn
A Visual Guide to Gemma 4 12Bnewsletter.maartengrootendorst.com2026-06-032026-10-07hn
Gemma 4 12B: A unified, encoder-free multimodal modelblog.google official2026-06-032026-10-07hn
Gemma 4 12B IT vs Gemma 4 E4B IT Comparison (2026) | LLMReferencellmreference.com2026-06-032026-10-07exa
Show HN: I made a Gemma 4 Mac app that names screenshots with local AIsnapname.app2026-05-312026-10-07hn
Show HN: Local LLM code-generation with Gemma 4 e2B via JSON AST to Clojuregithub.com2026-05-192026-10-07hn
Accelerating Gemma 4: faster inference with multi-token prediction draftersblog.google official2026-05-052026-10-07hn
Show HN: Prompt-to-Excalidraw demo with Gemma 4 E2B in the browser (3.1GB)teamchong.github.io2026-04-192026-10-07hn
Google Gemma 4 Runs Natively on iPhone with Full Offline AI Inferencegizmoweek.com2026-04-152026-10-07hn
Show HN: I benchmarked Gemma 4 E2B – the 2B model beat the 12B on multi-turnaiexplr.com2026-04-132026-10-07hn
I ran Gemma 4 as a local model in Codex CLIblog.danielvaughan.com2026-04-122026-10-07hn
A Visual Guide to Gemma 4newsletter.maartengrootendorst.com2026-04-092026-10-07hn
Show HN: Gemma 4 Multimodal Fine-Tuner for Apple Silicongithub.com2026-04-072026-10-07hn
Gemma 4 on iPhoneapps.apple.com2026-04-052026-10-07hn
Running Gemma 4 locally with LM Studio's new headless CLI and Claude Codeai.georgeliu.com2026-04-052026-10-07hn
April 2026 TLDR Setup for Ollama and Gemma 4 26B on a Mac minigist.github.com2026-04-032026-10-07hn
Google releases Gemma 4 open modelsdeepmind.google official2026-04-022026-10-07hn
Gemma 4: Byte for byte, the most capable open modelsblog.google official2026-04-022026-10-07hn

Papers

1 paper

Papers 1

arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).

Similar models

3 models

Similar models

  • Qwen2.5-Omni-7B · 10.7B
    same size band (L), same task (any-to-any), permissive license
  • MiniCPM-o-4_5 · 9.37B
    same size band (L), same task (any-to-any), permissive license
  • MiniCPM-o-2_6 · 8.67B
    same task (any-to-any), permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:34 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:01 UTC · Last verified 2026-10-25

Follow gemma-4-12B-it

Watch this model

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.