NVIDIA released three draft models to speed up Qwen3.8-27B. I converted them to GGUF and measured them on an RTX 4070 Ti SUPER with the same tokens as my previous post: on llama.cpp they speed things up by 14 % at most in aggregate, with sampling they only help on code and…
Badges
Badges come from GitHub once the author claims this feed.
Recent posts
every link opens the author’s post-
NVIDIA's draft models for Qwen3.8-27B on 16 GB: 43 tokens per second against 88 for MTP
-
An agent that reads my own posts and verifies the numbers aren't made up
Google released version 2.14 of its agent kit on November 4, 2026, and among other things graduated its local execution sandboxes out of experimental.
-
Cola2 on 16 GB: it fits with offload, scores 56% on ARC-Easy, but the paper's harness won't run
ByteDance Seed released Cola2 on November 6, 2026: a language model with continuous latents and a diffusion transformer in place of the autoregressive one.
-
Mage-VL-Cowork on 16 GB: it fits, it answers in 2.5 s, but it won't click
Microsoft released Mage-VL-Cowork, a 4-billion-parameter computer-use agent that scores 66% on OSWorld on paper.
-
The power grid jolts every minute: now we know why, and I simulated it
Power grid frequency dips for an instant exactly every 60 seconds across grids on several continents, and it was a mystery for years.
-
«SMS of Death»: an invisible text message can knock your Android off calls, even emergency ones
On 29 October 2026 a flaw was disclosed, nicknamed «SMS of Death» (CVE-2026-58882) and rated critical by Google: a special text message that never shows up in your inbox and hides who sent it can knock the call service off Android phones running 12 to 17, drop calls in progress…
-
Unlimited-OCR 2 on 16 GB: it reads as well as Datalab's Chandra 2, but it is not unlimited on my card
I tested Baidu's Unlimited-OCR 2, released on October 27, against Datalab's Chandra 2 and Marker, DeepSeek-OCR and Tesseract on an RTX 4070 Ti SUPER.
-
I measured Wild against mold again and mold won: the culprit was not the new versions, it was btrfs
In September Wild beat mold linking ripgrep.
-
Claude Code, Codex, OpenCode, OpenClaw, Hermes, mini-swe-agent and Raven on MiniMax-M3: 7 agent harnesses measured on Terminal-Bench 2.0
I had seven coding agent harnesses drive the same cheap model, MiniMax-M3, on 40 Terminal-Bench 2.0 tasks: 280 runs with every request saved and every pass reviewed by a blind judge.
-
Isolating a coding agent with gVisor or Kata: what it costs compared with Docker, measured
I measured what it costs to run a coding agent's work isolated in gVisor and in Kata Containers 4.3 (with QEMU and with Cloud Hypervisor) instead of a plain Docker container.
-
ComfyUI vs Forge on a 16 GB GPU: they tie on SDXL, and ComfyUI is twice as fast with Z-Image when the models do not fit
I measured ComfyUI, Forge Neo and the original Forge with the same files, seeds and settings on an RTX 4070 Ti SUPER, with diffusers as a baseline.
-
NVFP4 vs FP8 vs GGUF on a 16 GB GPU without Blackwell: NVFP4 generates 2.3 times faster than BF16 and ties with BF16 on the full 1319-question GSM8K
I measured the same NVIDIA model in BF16, NVFP4, MXFP8, FP8 and GGUF Q5_K_M on an RTX 4070 Ti SUPER.
-
NVFP4 without Blackwell: I ran NVIDIA's Qwen-Image on an RTX 4070 Ti SUPER and the NF4 grain was in 8 matrices
NVIDIA released Qwen-Image in NVFP4 for Blackwell only.