I happily picked up two used Nvidia Tesla P40 GPUs. I knew the P40 was old and ridiculously overbuilt. What I did not expect was how long the GPUs would become once the coolers were attached. Tried putting them in the case. Nope. The cooler end would not fit. I thought about it for about 10 seconds. Turns out the bottleneck was not the human. It was the case. So I grabbed a hammer and pliers and destroyed the front section of the case. After that both P40 GPUs fit perfectly. Put the cover back on and you cannot even tell from the outside. Apparently hardware fits if you hit it hard enough. Good experience.

Phonon 31m Gradium's open-weight 31m-parameter TTS runs offline on devices including a Raspberry Pi 5, with no GPU required. Gradium reports a 1.70% word error rate on 500 English sentences selected to challenge other models. gradium.ai/phonon

MiniMax M3.1 Flash: Affordable and Great at Search.

Why Chinese APIs Are Cheap

Ornith 1.5 122B. Stronger benchmarks and faster responses are nice. But running a 122B model on a single DGX Spark is what really caught my attention. huggingface.co/ornith-ai/Orn…

Rubick Rubick is an open-source gateway that serves chat, embeddings, decisions, and safety checks from one model instance in VRAM. Its README describes optional aLoRA adapters and KV-cache reuse, aiming to avoid separate models for each task. github.com/wfzyx/rubick

Amber-8 Amber-8 uses a local language model to interpret messy input, while deterministic code handles the rest of the program flow. Its examples compile to bytecode under 2 KB, and the project documents running it with a llama.cpp server. github.com/HuyNguyenAu/amber…

MiniMax M3.1 Flash is incredible. I wonder how many billions of parameters it has. Really curious. Is a revolution coming?

DeepSeek just open sourced DeepInfer. The inference engine behind its large scale production services. It supports DeepSeek V4.1 Flash with speculative decoding and optimized GPU execution. The full serving stack is not included yet. Still pretty interesting. github.com/deepseek-ai/DeepI…

Qwen3.8-27B-Coder390 A side-by-side RTX 5090 test found similar coding and math scores to the base model, with shorter reasoning, but weaker instruction following. The test used llama.cpp, Q6_K for both models and a 16k-token thinking budget. github.com/notwitcheer/llm-b…

vLLM 0.32.0 The release includes 814 commits from 363 contributors, with a major focus on lower-precision KV caches. For GLM-5.x on SM100, FP8 KV cache halves KV memory per rank; the release benchmark reports 2.3–5.5% more throughput at a 7–11% TTFT cost. github.com/vllm-project/vllm…

What is going on with MiniMax M3.1 Flash? It is performing way better in 3D than expected, and the pricing is insane.

Strata paired with two RTX 3090s is crazy good! Sounds like 3090 prices are going to go through the roof everywhere.

Gemini 4 Argon + Google ADK Google reports an 11 hour overnight run that cut long session in-memory session prep by about 40% and cold package imports by 21%. developers.googleblog.com/op…

MiniMax M3.1 Flash vs. DeepSeep V4.1 Flash 1017

Strata 0.2.0-beta Strata 0.2.0-beta brings an app, a Terminal interface and a separate inference engine. On a Tesla P100 16 GB, the release reports decode rising from 36.56 to 41.76 tok/s (+14.2%); its 10-prompt identity check returned identical answers. github.com/Niko1221/Strata/r…

MiniMax is back. It's great news, as a subscriber.

DeepSeek V4.1 Flash 1017 GGUF is here. Unsloth claims its 1 bit quantization retains around 88.2 percent accuracy while reducing the model from 508GB to 250GB. Local inference is possible with 128GB or more RAM. Strongest open model to date is a bold claim. Curious about the actual TPS and memory offloading overhead. huggingface.co/unsloth/DeepS…

MiniMax M3.1 Flash vs. DeepSeek V4.1 Flash 1017 xitter.cf/i/status/2112168890600…