OsamaJaber

4 ms·

1. 1.▲
   Agent loops stop re-reading your codebase on every turn(runinfra.ai)

   2 points by OsamaJaber 1mo ago|0 comments
2. 2.▲
   GLM 5.3 Flash faster and cheaper(runinfra.ai)

   1 points by OsamaJaber 1mo ago|0 comments
3. 3.▲
   The fastest and cheapest GLM 5.3 Flash endpoint(runinfra.ai)

   3 points by OsamaJaber 1mo ago|0 comments
4. 4.▲
   DeepSeek V4 Pro at 207 tok/s with the full 1M context, no quantization(runinfra.ai)

   3 points by OsamaJaber 2mo ago|0 comments
5. 5.▲
   Fastest Inference in MENA(runinfra.ai)

   2 points by OsamaJaber 2mo ago|0 comments
6. 6.▲
   by OsamaJaber 2mo ago

   soon :)
7. 7.▲
   by OsamaJaber 2mo ago

   you can give it a try
8. 8.▲
   DeepSeek V4 Flash at 278 tok/s, full precision, no quantization(runinfra.ai)

   10 points by OsamaJaber 2mo ago|4 comments
9. 9.▲
   The fastest full-precision Nemotron 3.5 Lightning endpoint (540 tok/s)(runinfra.ai)

   5 points by OsamaJaber 2mo ago|0 comments
10. 10.▲
    by OsamaJaber 2mo ago

    The comparison set is Gemma4-31B and Qwen3.6-27B, not the current Qwen Fair on size, but the headline numbers are against a model a generation back
11. 11.▲
    With software alone, one B200 beats the LPU and gets close to Cerebras(runinfra.ai)

    11 points by OsamaJaber 2mo ago|1 comments
12. 12.▲
    by OsamaJaber 2mo ago

    Expanding free access is mostly an inference cost Serving cheaply at that scale means routing, batching, and cache hits, not a better model :D
13. 13.▲
    Kimi K3 2.78T on One CPU with 8GB RAM(github.com)

    4 points by OsamaJaber 2mo ago|0 comments
14. 14.▲
    Mixture-of-Kittens: An MoE training megakernel for NVL72(twitter.com)

    3 points by OsamaJaber 2mo ago|0 comments
15. 15.▲
    Lossless Inference(runinfra.ai)

    4 points by OsamaJaber 2mo ago|0 comments
16. 16.▲
    DeepSeek V4 Flash 2.98x faster, lossless(runinfra.ai)

    1 points by OsamaJaber 2mo ago|1 comments
17. 17.▲
    DeepSeek-V4-Flash 2.98x faster on 4x B200, lossless(twitter.com)

    3 points by OsamaJaber 2mo ago|0 comments
18. 18.▲
    What LLM Inference Costs(twitter.com)

    2 points by OsamaJaber 2mo ago|0 comments
19. 19.▲
    Kimi k3 run on RTX 5090(github.com)

    4 points by OsamaJaber 2mo ago|0 comments
20. 20.▲
    Kimi k3 now runs on one consumer GPU(twitter.com)

    8 points by OsamaJaber 2mo ago|0 comments
21. 21.▲
    by OsamaJaber 3mo ago

    The gap I've hit generating GPU kernels with agents code that compiles and runs fine but is slower than the baseline. Validator says pass, result is useless. Speed targets have to be part of the check, not just correctness
22. 22.▲
    The AGI Compiler "Auto"(github.com)

    2 points by OsamaJaber 3mo ago|0 comments
23. 23.▲
    Compiler for LLMs, world models, and AGI(arxiv.org)

    2 points by OsamaJaber 3mo ago|0 comments
24. 24.▲
    Auto: The AGI Compiler(github.com)

    4 points by OsamaJaber 3mo ago|2 comments
25. 25.▲
    by OsamaJaber 3mo ago

    thread: https://x.com/Akashi203/status/2074495867449434389
26. 26.▲
    Build You Own Model(runinfra.ai)

    2 points by OsamaJaber 3mo ago|0 comments
27. 27.▲
    RunInfra: Optimize any open model down to the kernel, deploy in 5 min(runinfra.ai)

    2 points by OsamaJaber 3mo ago|0 comments
28. 28.▲
    Compiles any HuggingFace model into a single persistent megakernel(twitter.com)

    2 points by OsamaJaber 4mo ago|0 comments
29. 29.▲
    Mega Kernels, Written by Agents(arxiv.org)

    2 points by OsamaJaber 4mo ago|0 comments
30. 30.▲
    AutoMegaKernel: Compiling a LLM into a single CUDA kernel(arxiv.org)

    3 points by OsamaJaber 4mo ago|0 comments