What's new
By:

❓ Help May OpenClaw user ba dito?

Sort by date Sort by votes
Pahabol ko ito sa mga bagong gagamit ng openclaw or any claw apps. Be sure to choose models that have tool calling or function calling para di sumablay. Napansin ko lang at newbie rin he he.

Examples:
  • Claude 3.5 Sonnet: High success rate for tool calling and "Claw" logic.
  • GPT-4o: Reliable and smart with complex tools.
  • Gemini 1.5 Pro: Follows long instructions and uses tools effectively
  • Llama 3.1 (70B or 405B): Good tool calling support (available on Groq, Sambanova, and Nvidia).
  • GPT-4o-mini: Good at simple tool tasks.
  • Gemini 1.5 Flash: Fast and reliable for basic research.
Click to expand...
Kahit yung Claude Haiku from v3, Gemini Flash models from v1.5 pataas, pati Grok models supported yan. Check nyo sa kanilang site docs para sigurado yung tool calling modal nila, especially kung gagamit kayo ng local models na luma at maliit. Yung mga latest Qwen, Deepseek, Minimax, GLM, Kimi, Cogito (deepseek clone ng US) models..., supported na yan. Verify nyo muna pag may "Experimental" sa hulihan yung free model especially sa openrouter.
 
Upvote 0 Downvote
Draft said:
May free models din sa OpenRouter, ito mostly ang gamit ko.

You do not have permission to view the full content of this post. Log in or register now.
You do not have permission to view the full content of this post. Log in or register now.
Click to expand...
May naaala ako, at baka hindi alam ng iba na may pay-as-you-go bonus sa Openrouter. May free 1000 RPD pag nakagamit ka na ng +$10 credits. Keep that in mind. di ko lang alam kung hanggang saan yan tumagal - maybe after a year with no activity. Normal free credits is ~50 RPD (mataas up to 100 kung low-end models).
 
Upvote 0 Downvote
xLynx said:
yung sa groq.com paps, per token per AI ung limit. If reached na ang limit sa isang model, switch to other model na naman haha. Try ko subukan yung router na nirecommend mo. Thanks!
Click to expand...
...Pakner-in-crime, binalikan ko ulit yung mga dati kong ginamit na working pa rin, baka di mo pa na-test. Same principle sa ds-api. Pandagdag sa openclaw na token-eater he he.
Spoiler contents are visible only to Established Members.
 
Upvote 0 Downvote
bunay_o7 said:
oo upon setup pa pipiliin ka ng model , di advisable local llm di kakayanin ng Ordinary pc setup
Click to expand...
Tama. Nagkamali nga ako ng pagkaintindi noon at lately, kaya binalikan ko ulit kahit mahirap intindihin. Free online apis kasi gamit ko kaya hindi ko magamit ng matino siya noon from a single provider. Though pasok naman sa tokens/sec ng models, sa dami ng api calls ng openclaw, hindi na-meet yung required request per minute (rpm) ng tasks dahil sa capping ng free tier. Nagkaka-error pag tumigil yung api requests. Kaya tinigil ho na at inalam ko muna, para magawan ng paraan.

Usually, yung mga Tier 1 plan ng providers kaya yan. Pero sa free, i-compound mo pa sila para tumaas yung rpm na kailangan ng openclaw. Example, You do not have permission to view the full content of this post. Log in or register now.r. Para yang LLM manager that controls your assorted LLM api providers for the tasks. Kung hindi, sa light tasks lang pwede yang openclaw na parang chatbot or switch to any agentic AI assistant na pwede mong i-automate manually for basic needs, w/ light automation. Di biro gumamit ng openclaw at clones nya. Ang dilemma ng free api users ay RPM.

For chat and web search, ~10+ rpm is ok. Sa medium tasks like managing, cleaning, and reporting data, kailangan mo ng providers na may ~15 rpm+ (request/min) yung kayang ibigay, pero sa coding and heavy document analysis ay tataas hanggang 50 rpm or more, kaya recommended noon yung Claude 3.5 Sonnet as a benchmark to use openclaw na "Anthropic Tier 1 standard" ( na ~50 RPM yung capacity) - dahil dynamic ang api calls at token usages. Hindi ito fixed, kumbaga, kundi basehan lang para madaling mag-adjust yung user.

Ito naman yung sa LOCAL AI:

Estimated Desktop Requirements for Local Models : Sample only as actual will vary
To bypass the ~50 RPM requirement by using a local model, a desktop needs these specifications:
  • VRAM: A minimum of 12GB+ (e.g., RTX 3060/4060 Ti) is needed for a smooth experience with 7B or 8B models.
  • RAM: At least 32GB of system RAM is suggested if the GPU is not high-end.
  • Context Window: Configure the local provider to support at least 64k tokens. Otherwise, the agent may "forget" the beginning of the conversation during extended tasks.
Tingnan natin yung Logic and Effects: (Summarized by AI ito w/ minor editing )
The Core Dilemma: The KV Cache Tax
When running local AI, you are paying for two things in your VRAM:
  1. The Model File: An 8B model takes a flat ~5GB.
  2. The KV Cache (Context Memory): This grows dynamically. As you feed more tokens into the conversation, the KV cache expands, consuming more VRAM.
If Model + KV Cache + Operating System exceeds 12GB, the system offloads data to your 32GB System RAM. This keeps OpenClaw from crashing but significantly slows down performance.


The Master Configuration Matrix (For 12GB VRAM / 32GB RAM)


Context SettingStarting SpeedEnding Speed (At Max Context)Hardware Behavior (VRAM vs. RAM)OpenClaw Agent Consequence
8k Window~60 TPS~50 TPSFits 100% inside 12GB VRAM. Fast and cool.The "Goldfish" Effect: Too small. The agent will quickly forget its system rules, core memory, or previous code changes, causing loops or failures.
16k Window~60 TPS~35 TPSHugs the 12GB VRAM limit. Minor RAM spill if your PC background tasks spike.The Minimal Baseline: Works well for small scripts. Speed remains decent, but a single large code file can still push it to the memory limit.
32k Window~60 TPS~15–20 TPSThe Sweet Spot. Model fits in VRAM, but context memory s***** into System RAM at the end.Optimal Local Balance: OpenClaw has enough room to read 2–3 files and remember its instructions. Performance dips at the end, but the agent remains fully functional.
64k Window~60 TPS~3–8 TPSThe VRAM Wall. Massive context forces heavy reliance on slow System RAM.The Crawl: OpenClaw can handle huge codebases without forgetting anything, but it runs so slowly that background tasks and heartbeats may time out.


Critical Takeaways for a Local OpenClaw Setup

  1. The TPS is Dynamic: Your engine does not run at a single fixed speed. Every time OpenClaw sends a new request, the LLM must mathematically re-evaluate the entire history. Speed naturally scales down as the conversation length scales up.
  2. Memory Architecture Matters: Because you have 32GB of System RAM, your system acts as a safety net. You will not experience Out-of-Memory (OOM) crashes at 32k or 64k, but you will experience a processing bottleneck.
  3. The Recommended Compromise: For the smoothest local experience on a 12GB GPU, configure your local backend (Ollama/LM Studio) to 32k context (32768). It provides the best balance of required agent memory while keeping your execution speeds fast enough to handle tasks efficiently.
Click to expand...
Sa Local AI, sa TPS at hardware ka titirahin kahit unlimited yung RPM nya. Dahil habang tumataas na yung context window usage, bumababa yung tps ng dahan-dahan hanggang umabot sa dulo na ayaw nating mangyari. Either you use smaller quantized models para bumilis yung tps while making sure your hardware (VRAM and RAM) and kv cache can handle it. Adjust na lang sa context token window kung hanggang saan yung kaya ng tasks ng openclaw na gagamitin natin and the conditions mentioned earlier. Yan yung pagkaintindi ko ngayon.

Kaya dito sa sample specs ng Local AI, forced tayo na gumamit ng online apis for the heavy tasks para yung Local AI ay makikisali na lang sa small to medium tasks para di mapiga ng gusto at kung quality responses ang hanap, he he.

Note: Sa free online api users, alamin nyo lagi yung context window ng models na gagamitin ninyo dahil capped parati yan compared sa ρáíd model. Ang example nyan is Gemini na capped sa 32k. Pero nagamit ko siya sa coding project without problems since console app lang yon sa opencode.
 
Upvote 0 Downvote
Draft said:
Fixed, ayaw gumana ng OpenROuter option, pero working fine with OpenAI.
Click to expand...
Buti't napagana mo. Mas bagay si OC sa prime models he he. Yang OpenAI, 500 rpm na ang bigay ngayon kaya no problemo. Nasa tier plan ng big providers yung kailangan ni OC. Paningit yung free models to save costs.
Sa tulad kong homeboy lang, stable na ako without openclaw. Pero ang dami kong natutunan at nadiskubre dahil sa thread na ito. One thing leads to another!
Sana, di ka mahirapan sa token eater na yan. Just use the token saving apps.
 
Upvote 0 Downvote
Kung gusto nyong sumubok ng Openclaw, mas ok kung kabisaduhin nyo yung Claude Code muna or any similar cli app para maganda yung progression ng transfer, especially using AI Agents. Openclaw directly is problematic for new users. Yung familiarity sa gamit ng CLAUDE.md, AGENT.md, SKILLL.md, will be much easier.

Since openclaw thread ito. might as well treat you with a free option—no harm in trying as it may still work or not!
Spoiler contents are visible only to Established Members.
Skeletonize nyo na lang kung ayaw nyo sa Openclaw, he he.
Hindi ito for production, but for personal use only. Visit the issues section for further details.

Pero sa'kin, mamili kayo ng models na pasok sa inyong need for openclaw, na para kayong may local AI. Pili kayo ng models na +30rpm, na may libong RPD or decent free tokens/day; buhay na kayo. Legit providers yan, hindi reverse-engineered web-to-API tulad noong una na pwedeng ma-ρá†ch anytime.
Spoiler contents are visible only to Established Members.
 
Upvote 0 Downvote
xLynx said:
View attachment 4177166
di ko mahanap language settings ahahah
Click to expand...
Huli na si free-ds pakner. Lahat ng devs, nag-abandon na. Either buhay yung api., litaw or live yung models, depende sa dev na ginamit mo, pero 100% banned yung accounts pag-load ng api calls he he. Waiting for new updates.
GPT_API_free at gpt4free na lang yung matatag na 2 - 3 yrs kong gamit (5 - 30 - +200 RPD) na pandagdag sa routers for OC at cli-agents.
Ganyan talaga pag api hunter, weather-weather lang. Pero nasa WPS area sila - mailap, makunat, pero pwedeng ipunin yung barya na $ to unlimited kahit foreigner ka. Sa RE-method, di lahat makakahabol dahil hindi well-documented. Oras lang, baka patay na. másáráp subakan for the adrenaline/dopamine rush na hanap ko.
 
Upvote 0 Downvote
Write a reply…

Similar threads

About this Thread

  • 73
    Replies
  • 2K
    Views
  • 19
    Participants
Last reply from:
fleu9586

New Topics

Trending Topics

Online now

Members online
880
Guests online
702
Total visitors
1,582

Forum statistics

Threads
2,276,945
Posts
28,973,238
Members
1,229,655
Latest member
Zfkirke0109
Back
Top