Erwan Novianto
58 posts
dahsyat ini keliatannya Kimi K3, yang sebelumnya 2.7 code aja udah keren, cuma madalah harganya aja ini moga2 bisa murah ya
Kimi.ai@Kimi_Moonshot
Introducing Kimi K3: Open Frontier Intelligence 🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal 🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts 🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost 🔹 Built for long-horizon agentic coding and self-evolving workflows Kimi K3 is now live on on Kimi.com, Kimi Work, Kimi Code, and the Kimi API. Open Weights by July 27, 2026. 🔗 API: platform.kimi.ai 🔗 Tech blog: kimi.com/blog/kimi-k3
Indonesia
@AuliaBismar @airplanestar_ Wah coba dulu, sy jg awalnya gak minat jg krn biasa pake opencode + omo, tp skr kok enak ya, beda project cukup 1 window
Indonesia
@airplanestar_ Gak ada versi cli? Gak minat lagi pake beginian
Indonesia
gak nyangka Zcode se smooth ini, selain bisa pasang custom provider tanpa ribet, doi terus terusan di update sama dev nya, penggunaan tiap hari makin betah keren.
pake Zcode sangat nyaman dengan model model dari negri tirai bambu, agentic sangat terasa, kureng nya kalo udah ngecompress context windows, lama banget, mungkin karna bandwith internet gue 40 MB
selain codex, antigravity, lu pake apa lagi buat ngoding?
Indonesia
@airplanestar_ Eh sama sy jg lg ketagihan pake zcode krn pake glm nyaris ga abis2
Indonesia
Hasil trainingan2 distilation seperti ini kenapa selalu lebih turun dari base ya? Base-nya aja dapet 94%
Hugging Models@HuggingModels
Gemma 4 12B Coder is here and it's a game changer for local code generation. This GGUF model packs Google's latest gemma-4 architecture into a compact 12B size, perfect for running on consumer hardware. It's optimized for reasoning and thinking, making it ideal for developers who want fast, private coding assistance without the cloud.
Indonesia
@CommandCodeAI I'm building a sales AI agent with @CommandCodeAI
English
We're close to hitting 10K on X. Should celebrate.
Giving away 3 Pro subs
STEPS
1. Follow @CommandCodeAI
2. Repost this post
3. Reply with what you're building with Command Code
We'll pick randomly soon as we hit 10K. LFG!
English
@remotesimple @deepseek_ai @Alibaba_Qwen @Zai_org GLM 5.2 dulu klo mentok coba Minimax m3 , klo mentok juga Kimi 2.7 code, klo mentok juga ke qwen 3.7, semua setara sih cuma kadang masing2 punya sudut pandang beda2 dan bisa saling liat kelemahan yg lain
Indonesia
gue Udah nyobain semuanya Model.
Beware Fomo Vibe Post (Hanya menurut pengalaman saja)
Ai China yang Worth untuk sekarang cuman
- DeepSeek V4 Pro @deepseek_ai (murah dan sangat worth it)
- Qwen 3.7 @Alibaba_Qwen (mahal tapi emang sejago itu)
- GLM 5.2 @Zai_org (biasa aja dan server nya sampah, sering timeout)
ingat ini hanya Pengalaman gue. silahkan dicoba lagi siapa tahu ternyata gue gak bisa memaksimalkan Modelnya.
Ronin@DeRonin_
My entire AI stack is now Chinese 🇨🇳 87% cheaper. same revenue swaps by task: 1. reasoning / backend brain Opus 4.8 → Kimi K2.7 benchmark gap: ~8% · price: ~11x cheaper 2. code generation GPT-5.5 → Qwen 3.7 Max benchmark gap: ~18% · price: ~7x cheaper 3. agent loops + tool calling Sonnet 4.7 → GLM 5.2 benchmark gap: ~3% · price: ~5x cheaper on input 4. cheap volume / bulk processing GPT-5.5 mini → MiMo V2.5 benchmark gap: ~6% · price: ~12x cheaper 5. image generation GPT-Image-2 → Wan 2.5 benchmark gap: ~5% · price: ~8x cheaper 6. video generation Sora 2 → Kling 3.0 benchmark gap: roughly equal · price: ~6x cheaper [ result after 30 days: ] operating costs dropped 87%, output quality dropped 4% on average, revenue unchanged the most important that these models will be not banned in a month and i can run them locally nobody will steal my data and i can learn them as i need full article drops tomorrow with: > exact routing logic per task type > the 2 cases where I still pay for American > the migration playbook anyone can copy in a weekend VERY IMPORTANT to get migrated now, while it's not too late
Indonesia
おぉぉ！VRAM16GBでGLM-5.2が動いた！
RDIMMer勝利w
日本語
Masalahnya bukan mertua yang salah. Bukan IG yang jahat. Bukan ChatGPT yang bohong.
Masalahnya:
Kita gak tahu kapan harus tanya siapa.
Kita pakai sumber hiburan untuk keputusan medis. Kita pakai sumber emosional untuk keputusan klinis.
Dan anak kita yang menanggung akibatnya.
Indonesia
Anaknya demam 38,5°C. Jam 10 malam.
Dia tanya 4 sumber. Dapat 4 jawaban.
Mertua: "Kerokan. Angin masuk."
IG: "Bawang merah parut di ketiak. Alami."
ChatGPT: "Kompres hangat, pantau 24 jam."
Dokter: "Parasetamol sesuai berat badan. Kalau kejang, langsung IGD."
Dia bingung. Dia pilih yang paling banyak di-like di IG.
Dan itu keputusan yang hampir dia sesali seumur hidup.
Indonesia
@perogeremmer Saya jg pake opencode go $10, tp utk ds4f dan mimo v2.5 ada free dari opencode zen, jd sy pake yg free gak abis2 limitnya, klo mentok baru pake minimax m3 dari opencode go
Indonesia
Saya kasih tipsnya biar kalian hemat.
Opencode Go - $10
Dorong pake model Deepseek v4 flash atau Mimo v2.5 flash lebih banyak, trust me, in simple case dia udah cukup helpful. Apalagi buat case mahasiswa dan fresh grad.
Kalo bener-bener yang harus mikir, naikin variant ke high atau max, atau switch ke Deepseek v4 atau Mimo v2.5 pro.
Fallback, topup 5$ di Deepseek API langsung, kalo misal dah habis yang opencode go, pake fallback dan fokus ke deepseek v4. Ini saya pake buat AionUI, so far 5$ belom habis karena ya saya pake flash mulu.
Command Code - 1$
Lagi-lagi dorong pake Mimo dan Deepsek v4 flash. Udah worth it banget walau saya gak suka sih CLI macam gini, lebih suka OpenCode.
Xiaomi Mimo Token Plan - 6$
Lumayan bisa multimodal, dan Mimo v2.5 tuh udah bagus buat coding.
---
Kuncinya di bagaimana lu bisa jelasin dengan baik ke AI, dan gak ngarep AI tahu isi kepala lu.
So far saya gapernah menghabiskan 1 juta rupiah untuk ngoding, dan tetap bisa produktif kok 😁
satyaxbt@satyaXBT
mau tanya dong, subscription AI atau bayar API kalian nyampe berapa dollar perbulan?
Indonesia
Dear Algoritma X,
Kakak-kakak udah pada makan malam belum? Mau ini gak?
Indonesia
Erwan Novianto retweetledi
Thank you all for the interest in my setup! 🙏
I'll read and reply to every message.
As promised:
Tech specs of my AI cluster 👇
- Board: ROMED8-2T
- CPU: EPYC 7302
- RAM: 256GB DDR4-3200
- Storage: 2TB NVMe
- Risers: ADT-Link MCIO
- PSU: 2xSuper Flower Leadex 2000W
Network:
- ConnectX-6
- SB7800 InfiniBand
Cooling:
- 8x12mm fans exhausting GPU heat
- 9733 for the ConnectX-6
Max Zanoga@zanoga
Finally finished building my AI datacenter! 🚀 32x3090s across 4 servers (8 GPUs each), all connected over InfiniBand. The whole setup is solar-powered with a massive battery bank and generator backup. More technical details and benchmarks coming soon.
English
Erwan Novianto retweetledi
Grok Build + local models is stupidly easy.
Just add this to ~/.grok/config.toml and you’re done:
[model.deepseek]
model = "deepseek-v4-flash"
base_url = "http://localhost:8000/v1"
name = "deepseek-v4-flash"
Replace name/model & ip/port as needed.
Then run grok as usual 👇
English
Erwan Novianto retweetledi
2/ KESALAHAN TEPUNG
Hampir semua orang pakai 100% tepung terigu.
Logikanya masuk akal. Terigu = kriuk.
Tapi ini jebakan.
Terigu itu tinggi gluten. Dan gluten sifatnya rakus nyerap uap air. Hasilnya? Garing 5 menit, lembek 10 menit.
Formula yang bener:
→ 70% terigu
→ 25% tepung beras (kriuk ringan)
→ 5% tapioka (ngunci kriuknya biar tahan lama)
Indonesia
Erwan Novianto retweetledi
Saya bagi model GRATIS lewat vikey.ai
Anggep aja bagi-bagi token ya 🤣
Daftar di vikey.ai (cepet, tanpa verif email/kartu)
Login -> menu API Keys -> Create New Keys -> copy
Setting di agent kamu:
host: api.vikey.ai/v1
model: vikey/vclaw
apikey: dapetin ini di webnya vikey
Gratisnya khusus model vikey/vclaw & ini yg gw pake harian.
Cukup Lah
Indonesia
Indonesia
Halo! Sayangnya **tidak ada cara** upgrade RAM di Mac Studio M4 Max setelah dibeli. Unified memory-nya disolder langsung ke logic board (bukan slot seperti PC).
Kamu cuma bisa pilih konfigurasi RAM (36GB/64GB/128GB) **saat beli** dari Apple. Setelah itu fixed.
Berbeda dengan storage Mac Mini M4 yang pakai modul proprietary & bisa di-swap pakai third-party (~$300). RAM Mac Studio beda cerita total.
Pilih spek RAM yang pas dari awal ya!
Indonesia
The claim is mostly accurate with caveats. Apple's M4 Mac Mini storage upgrades are heavily marked up (~$800 extra for 2TB vs ~$599 base 256GB model).
The SSD is a removable proprietary module (controller lives in the M4 SoC; not soldered like older Macs). Third-party compatible modules (~$270–$350 from Expand Mac Mini and similar) deliver full native speeds.
DIY swap is real and documented (iFixit guide exists), but:
- Not a standard M.2
- Real cost closer to $300 than $100-150
- Full process (backup + possible DFU restore) exceeds 15 minutes
- Voids warranty; use reputable kits only
Big savings if you're handy and back everything up first. Proceed at your own risk.
English
Mantap lokal API inference, NaraRouter, free 5M token per day, valid, dan tested.
Aku kasih trik topup aja 20rb biar bisa akses model yang lebih besar, kalau tanpa topup hanya beberapa model terbatas, bisa diliat ini list
gas langsung daftar di sini (mengandung reff boleh dihapus): router.naraya.ai/register?ref=Z…
enaknya disini harganya pakai rupiah dan bisa pake QRIS. Oh ya yang butuh setting di opencode nya silakan request
Indonesia
Erwan Novianto retweetledi
Proud to introduce a new 27B post-trained model
After being impressed by both Fable and Kimi 2.7 Coder, I wanted to see how much of the positives i could distill from both, and I think you all will be pleased!
Links below!
English