What am I working on? Condensing everything I do into one place: – local AI / LLMs
– inference + benchmarking
– hardware + cluster builds
– LLM research + notes
– agent workflows
– real-world perf (tokens/sec, concurrency, thermals) All into a single, searchable, indexable
@theahmadosman
-
Condensing local AI and LLM work into one searchable space
By
–
-
An agent web stack using SearXNG, Firecrawl, and Camofox
By
–
PRO TIP My Agent Web Stack – SearXNG: candidate source discovery
– Firecrawl: known-URL scrape and crawl
– Camofox: browser fallback for JS/interaction Search -> Extract -> Interact P.S. Give this to your favorite agent & tell it to setup these tools to use with local models -
Qwen 3.5’s overthinking helps infer intent, boosting Gemma 4’s performance
By
–
Qwen 3.5 tends to overthink, which ironically helps it better infer intent Gemma 4 is the opposite, you have to spell everything out (probably guardrails) Had Qwen rewrite my prompts, then used those on both models Gemma’s performance jumped noticeably
-
People are sleeping on ExLlamaV3 inference engine potential
By
–
people are sleeping on exllamav3 there's so much to do with, and learn from, that inference engine
-
Ahmed Osman’s Top Open Source and Home LLMs
By
–
> Best overall opensource LLM is GLM-5.1
— Ahmad (@TheAhmadOsman) 13 avril 2026
> Best openweight model to run at home is MiniMax M2.7 https://t.co/l9Sd4Oy6XT> Best overall opensource LLM is GLM-5.1 > Best openweight model to run at home is MiniMax M2.7
-
Context length benchmarks up to 180k tokens
By
–
Ran some benchmarks on different context lengths all the way to 180k Findings below
-
Exploring dense new format with tests in BF16, FP8, NVFP4 and TP scaling
By
–
it's very dense, a new format that I am playing with still need to test it in concurrency + try it in nvfp4 hopefully will be able to compare perplexity in bf16 / fp8 / nvfp4 as well as TP performance jumps from 2 -> 4 -> 8 nodes across all 3 formats most important thing
-
Testing concurrency, nvfp4, and TP performance across BF16, FP8, nvfp4
By
–
still need to test it in concurrency + try it in nvfp4 hopefully will be able to compare perplexity in bf16 / fp8 / nvfp4 as well as tp performance jumps from 2 -> 4 -> 8 nodes across all 3
-
MiniMax M2.7 on 4x DGX Sparks with vLLM and OpenCode
By
–
MiniMax M2.7 at home running on 4x DGX Sparks
— Ahmad (@TheAhmadOsman) 12 avril 2026
vLLM serving full BF16 weights, 200k context
OpenCode having the model monitor its own hardware and report thermals, tokens/sec, TTFT, and other runtime stats in real time
What benchmarks / workflows / things do you wanna see next? https://t.co/iCR8UsLs8E pic.twitter.com/I04c7j6UkgMiniMax M2.7 at home running on 4x DGX Sparks vLLM serving full BF16 weights, 200k context OpenCode having the model monitor its own hardware and report thermals, tokens/sec, TTFT, and other runtime stats in real time What benchmarks / workflows / things do you wanna see next?
-
Parallel Agentic Workflows skill now live for all CLI agents
By
–
Skill for Parallel Agentic Workflows is now live Works w/ any CLI agent harness
(Codex, Claude, Kimi, OpenCode, Droid, etc) Be warned this was vibecoded from my workflows, not fully tested Should be a GREAT STARTING POINT nevertheless Give the screenshot below to your agent