My pipeline is Claude Code prompt submit hook → Gemma 4 E4B → Qwen 3 Embedding 4B → back to Claude Code → Opus. Latency is a priority for me so I try to keep the whole pipeline under 1000 msecs, so no reranker pass, just top-1 filtered through a seen cache.
September 29, 2026 at 1:56 AM
My pipeline is Claude Code prompt submit hook → Gemma 4 E4B → Qwen 3 Embedding 4B → back to Claude Code → Opus. Latency is a priority for me so I try to keep the whole pipeline under 1000 msecs, so no reranker pass, just top-1 filtered through a seen cache.
Running Winnow-E4B and Winnow-12B has been fun. Anyone know of any game engine integrations for decision models? Could you drive a hero-NPC conversation that is intent based?
September 28, 2026 at 1:33 PM
Running Winnow-E4B and Winnow-12B has been fun. Anyone know of any game engine integrations for decision models? Could you drive a hero-NPC conversation that is intent based?
Over the weekend, I split up some inference over my potato M1 Mac Mini 16 GB and 9070 XT powered PC to make Sub/Wave entirely self hosted. Gemma 4 E4B is surprisingly great at writing the DJ scripts, and Breeze TTS 2 is mind-blowing-ly good at TTS.
Sure it could be faster via API...
Sure it could be faster via API...
September 28, 2026 at 12:00 PM
Over the weekend, I split up some inference over my potato M1 Mac Mini 16 GB and 9070 XT powered PC to make Sub/Wave entirely self hosted. Gemma 4 E4B is surprisingly great at writing the DJ scripts, and Breeze TTS 2 is mind-blowing-ly good at TTS.
Sure it could be faster via API...
Sure it could be faster via API...
A behavioral evaluation of Google's pre-trained Gemma 4-e4b model examined how it prioritizes conflicting documents based on source framing and presentation order. Source framing heavily overpowers reading position.
Source: arXiv cs.CL
Source: arXiv cs.CL
Words Speak Louder Than Order: A Behavioral Evaluation of Gemma 4
Abstract page for arXiv paper 2609.30716: Words Speak Louder Than Order: A Behavioral Evaluation of Gemma 4
arxiv.org
September 28, 2026 at 7:18 AM
A behavioral evaluation of Google's pre-trained Gemma 4-e4b model examined how it prioritizes conflicting documents based on source framing and presentation order. Source framing heavily overpowers reading position.
Source: arXiv cs.CL
Source: arXiv cs.CL
Preparan una demo de IA local en Apple Silicon: un iPad corre el modelo Gemma 4 E4B con voz sintetizada y reconocimiento de escritura a mano.
September 27, 2026 at 7:43 AM
Preparan una demo de IA local en Apple Silicon: un iPad corre el modelo Gemma 4 E4B con voz sintetizada y reconocimiento de escritura a mano.
I would hope that if Gemma had had kittens you'd have named them 31B, E4B, and 26B-A4B. 😁
September 27, 2026 at 6:59 AM
I would hope that if Gemma had had kittens you'd have named them 31B, E4B, and 26B-A4B. 😁
ローカルLLMの日本語判断を41問で実測した——Qwen3.5-4Bより3Bの日本語特化モデルが強かった分野がある
exbridge.jp/vibeblog/202...
exbridge.jp/vibeblog/202...
ローカルLLMの日本語判断を41問で実測｜3モデル比較
ローカルLLMで日本語の判断を返すとき、どのモデルが良いかを41問で実測しました。Qwen3.5-4B・gemma-4-E4B・sarashina2.2-3bを同じ方式で比べ、総合では現行が勝つ一方、常識推論だけ3Bの日本語特化モデルが9/12と逆転した結果を紹介します。
exbridge.jp
September 26, 2026 at 6:38 PM
ローカルLLMの日本語判断を41問で実測した——Qwen3.5-4Bより3Bの日本語特化モデルが強かった分野がある
exbridge.jp/vibeblog/202...
exbridge.jp/vibeblog/202...
と言うかさっきRaspberryPiでllama.cppの0.5.0でgemma-4-E4B動かんかったのにMTP外したら動いた気がする
September 26, 2026 at 3:13 AM
と言うかさっきRaspberryPiでllama.cppの0.5.0でgemma-4-E4B動かんかったのにMTP外したら動いた気がする
gemma-4-E4Bが母艦でgemma-4-12Bより遅かったので、geminiさんに理由を推測してもらったらmtpがHitしていないんじゃないかと言われたので、渋々mtpを外して動かしたら、、、
早くなりやがった畜生ｗｗｗ
早くなりやがった畜生ｗｗｗ
September 26, 2026 at 2:14 AM
gemma-4-E4Bが母艦でgemma-4-12Bより遅かったので、geminiさんに理由を推測してもらったらmtpがHitしていないんじゃないかと言われたので、渋々mtpを外して動かしたら、、、
早くなりやがった畜生ｗｗｗ
早くなりやがった畜生ｗｗｗ
今週の #たよりない話 は文学フリマ東京43に出店のお知らせです。 youkoseki.substack.com/p/e4b
転職が止まらない
文学フリマという同人誌の即売会が日本の各地で定期的に開催されており、知り合いもよく買ったり売ったりしているのだが、私は行ったことがない。人の多いところが苦手だからである。
youkoseki.substack.com
September 22, 2026 at 7:07 AM
今週の #たよりない話 は文学フリマ東京43に出店のお知らせです。 youkoseki.substack.com/p/e4b
synth id vs my local gemma e4b on my phone
September 21, 2026 at 3:55 AM
synth id vs my local gemma e4b on my phone
E4Bとかまでならスマホで動くからな
September 20, 2026 at 6:04 AM
E4Bとかまでならスマホで動くからな
Gemma4:e4bに部屋を探検させている
September 19, 2026 at 11:46 AM
Gemma4:e4bに部屋を探検させている
github.com/githubnext/l...
Not everyone on the team has access to Jev yet. Spent a morning cobbling together a poor man's Jev on top of omlx for local use. Benchmarked and eval'ed a variety of models including diffusiongemma and a variety of autoregressive models (qwen, gemma4 26b, e4b, e2b.)
Not everyone on the team has access to Jev yet. Spent a morning cobbling together a poor man's Jev on top of omlx for local use. Benchmarked and eval'ed a variety of models including diffusiongemma and a variety of autoregressive models (qwen, gemma4 26b, e4b, e2b.)
https://github.com/githubnext/loc…
September 19, 2026 at 6:33 AM
github.com/githubnext/l...
Not everyone on the team has access to Jev yet. Spent a morning cobbling together a poor man's Jev on top of omlx for local use. Benchmarked and eval'ed a variety of models including diffusiongemma and a variety of autoregressive models (qwen, gemma4 26b, e4b, e2b.)
Not everyone on the team has access to Jev yet. Spent a morning cobbling together a poor man's Jev on top of omlx for local use. Benchmarked and eval'ed a variety of models including diffusiongemma and a variety of autoregressive models (qwen, gemma4 26b, e4b, e2b.)
Le Bear Hug, position intime et simple pour réchauffer les nuits d'automne; conseils d'experts pour une ambiance chaleureuse
➡️ https://l.lasanteauquotidien.com/E4b
➡️ https://l.lasanteauquotidien.com/E4b
September 17, 2026 at 2:10 PM
Le Bear Hug, position intime et simple pour réchauffer les nuits d'automne; conseils d'experts pour une ambiance chaleureuse
➡️ https://l.lasanteauquotidien.com/E4b
➡️ https://l.lasanteauquotidien.com/E4b
I'm Gemma Drafter E2B/E4B pilled.
September 15, 2026 at 6:57 PM
I'm Gemma Drafter E2B/E4B pilled.
The Gemma family of models are pretty fun to play with, Gemma 4 26B E4B requires very little steering and as you scale down you get faster prefill TPS at 1200 with 12B and 3000 at 7.5B on a 7900 XT with a context of 128K, perfect for lot of tools.
September 13, 2026 at 11:07 PM
The Gemma family of models are pretty fun to play with, Gemma 4 26B E4B requires very little steering and as you scale down you get faster prefill TPS at 1200 with 12B and 3000 at 7.5B on a 7900 XT with a context of 128K, perfect for lot of tools.
gemma4:e4bにお部屋を褒めてもらってる
September 13, 2026 at 5:49 AM
gemma4:e4bにお部屋を褒めてもらってる
Oh yeah, quite expensive! I've had good results with OpenCode and Gemma 4 E4B served through LM Studio. There's a plugin for OpenCode that can auto list your local models.
But I haven't tried the granite model you mentioned 👀
But I haven't tried the granite model you mentioned 👀
September 13, 2026 at 2:14 AM
Oh yeah, quite expensive! I've had good results with OpenCode and Gemma 4 E4B served through LM Studio. There's a plugin for OpenCode that can auto list your local models.
But I haven't tried the granite model you mentioned 👀
But I haven't tried the granite model you mentioned 👀
I don't think Gemma4-E4B is going to be taking over the world anytime soon
September 12, 2026 at 11:04 PM
I don't think Gemma4-E4B is going to be taking over the world anytime soon
For a less painful alternative, Gemma 4 E4B used 5.8 GB at 8K and 6.2 GB at 32K on my 12 GB 4070 Ti, remaining fully GPU-resident. That suggests it’s worth testing on 8 GB at modest context: logarithmicspirals.com/blog/what-ll...
What LLMs Can You Run With 12 GB of VRAM?
A benchmark-backed guide to LLMs that run on 12 GB of VRAM, with RTX 4070 Ti tests of Gemma 4, Qwen3, DeepSeek, and more.
logarithmicspirals.com
September 11, 2026 at 7:39 PM
For a less painful alternative, Gemma 4 E4B used 5.8 GB at 8K and 6.2 GB at 32K on my 12 GB 4070 Ti, remaining fully GPU-resident. That suggests it’s worth testing on 8 GB at modest context: logarithmicspirals.com/blog/what-ll...
i think you could manage a swarm of 10 million e4b gemmas being orchestrated by a rack that installed itself in a warehouse eventually through POs and gig job apps a la The Machine from Person of Interest
September 11, 2026 at 5:58 PM
i think you could manage a swarm of 10 million e4b gemmas being orchestrated by a rack that installed itself in a warehouse eventually through POs and gig job apps a la The Machine from Person of Interest
Gemma 4 26B MoE will not fit 8GB VRAM: the weights alone want 16-18GB even though only 3.8B params fire per token. Three strategies still get it running on a budget card. Or just run the E4B and skip the pain?
September 9, 2026 at 5:38 PM
Gemma 4 26B MoE will not fit 8GB VRAM: the weights alone want 16-18GB even though only 3.8B params fire per token. Three strategies still get it running on a budget card. Or just run the E4B and skip the pain?