OsamaJaber
Free Cached Input
Ornith 1.5 122B at 302tps
The fastest DeepSeek v4.1 flash 1017 – Day 0
Is this the best inference for agents?
DeepSeek v4.1 Flash at 378 tok/s 99.7% Cache hit rate
GLM 5.3 flash on AMD GPUs 670 tok/s
Agent loops stop re-reading your codebase on every turn
GLM 5.3 Flash faster and cheaper
The fastest and cheapest GLM 5.3 Flash endpoint
DeepSeek V4 Pro at 207 tok/s with the full 1M context, no quantization
Fastest Inference in MENA
Re:DeepSeek V4 Flash at 278 tok/s, full precision, no quantization
on Sunday August 16, 2026 @12:30PM (#49321492)
soon :)
Re:DeepSeek V4 Flash at 278 tok/s, full precision, no quantization
on Saturday August 15, 2026 @02:54PM (#49313234)
you can give it a try
DeepSeek V4 Flash at 278 tok/s, full precision, no quantization
The fastest full-precision Nemotron 3.5 Lightning endpoint (540 tok/s)
Re:Meta Muse Glimmer – open weights 30B local coding model
on Monday August 10, 2026 @08:57AM (#49243020)
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
With software alone, one B200 beats the LPU and gets close to Cerebras
Re:Improving GPT-5.6 Sol in ChatGPT—and expanding access for free users
on Thursday August 06, 2026 @04:39PM (#49202206)
Expanding free access is mostly an inference cost
Serving cheaply at that scale means routing, batching, and cache hits, not a better model :D
Kimi K3 2.78T on One CPU with 8GB RAM
Mixture-of-Kittens: An MoE training megakernel for NVL72
Lossless Inference
DeepSeek V4 Flash 2.98x faster, lossless
DeepSeek-V4-Flash 2.98x faster on 4x B200, lossless
GLM-5.3 Soon
$0.09 and $290.12 are both the price of 1M output tokens
What LLM Inference Costs
Kimi k3 run on RTX 5090
Kimi k3 now runs on one consumer GPU
Re:DSLs Enable Reliable Use of LLMs
on Wednesday July 15, 2026 @08:18AM (#48919685)
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