NockAI @useNockAI

Small open models get judged on benchmarks, but a lot of the gap is in how you ask. Same 4-bit model, same frames, one changed sentence: 3 of 12 became 10 of 12 on our screen recording questions. Measure the cheap fix before you train.

Overall it went from 31 to 35 of 56. Not free though: sport got worse, 4 to 1 of 8. Twice as many frames instead of a better prompt: 34 of 56. So the new prompt stays, and training goes after what it can't fix.

Example: when does the app switch to dark mode? The true answer is 26 to 34 s. Old prompt: 26-27 New prompt: 26-40 The change: ask for the whole stretch, from the first frame where it starts to the last frame where it's still visible.

We didn't train anything. One sentence in the prompt took "find the moment" on short screen recordings from 3 of 12 to 10 of 12. The model already knew when things started. It just stopped too early.

Answer to yesterday's clip: the error shows from about 34 s to 42 s. Qwen3.5 4B said 34-42. Gemma 4 E4B said 34.5-39.9. SmolVLM2 2.2B said 29. That one was easy. Short screen recordings overall were not: 3 of 12 for the best model.

NockAI @useNockAI

2 days ago

In this clip, when does the error message show up? Our next model has to answer questions like this about your own videos, offline. We wrote 123 of them before running a single model.

2 1 10 413 0

Visual understanding in a small Qwen3.5 build is the part we care about most. On our video test, Qwen3.5 4B at 4-bit in llama.cpp kept almost everything the full model scored: 31 vs 34 of 56 moments, 56 vs 57 of 60 questions. Small builds hold up better than you'd think.

PrismML @PrismML

3 days ago

Today, we’re releasing Bonsai 2 2B, 4B, and 9B. Based on Qwen3.5, these new ternary models bring stronger reasoning, visual understanding, and tool calling to compact footprints designed for everyday devices: • 2B: 0.43GB • 4B: 0.95GB • 9B: 2.0GB For this release, we

143 384 3K 307K 2K

Finding the moment is the hard part, and it matches what we just measured. On our 20 video test, the best small open base answered 57 of 60 questions about what's on screen, but found the right moment only 34 of 56 times. Short screen recordings were the worst: 3 of 12.

Aran Komatsuzaki @arankomatsuzaki

4 days ago

VideoJev: System One Model for Video Understanding and Interaction - >100x QA speedup vs. Thinking - +6.7 / +10.8 pts mIoU gain over 4B / MoE backbones - Opensourced 4B and 35B variants

9 56 389 25K 351

Asking about a video already works. Finding the moment doesn't yet: on short screen recordings the best one got 3 of 12. The 4-bit build in llama.cpp kept almost all of it: 31/56, 56/60, 6/7. So Qwen3.5 4B is the base, and finding the moment is what we train next.

First run, three open bases, same frames, same prompts: Qwen3.5 4B from @Alibaba_Qwen: 34/56 moments, 57/60 questions, 7/7 summaries Gemma 4 E4B from @googlegemma: 19/56, 56/60, 6/7 SmolVLM2 2.2B from @huggingface: 1/56, 37/60, 1/7

Our next model answers questions about your own videos, offline. Family clips, meetings, screen recordings. Before training it, we built the test: 20 videos, 54 minutes 56 “when does it happen?” questions 60 questions about what’s on screen 7 summaries

In this clip, when does the error message show up? Our next model has to answer questions like this about your own videos, offline. We wrote 123 of them before running a single model.

Mostly what the editor hands it: your selection or the open file. It’s built for the quick single answer (a function, a regex, a fix in one place), not for reasoning across a codebase. For multi-file work we just released Nock Agent 31B: it reads the project with tools, runs the tests, then edits. Bigger model, different job. huggingface.co/Nock-AI/nock-a…

Nock Coder now works inside @code. Install the @ollama extension, open Chat, pick nockai/nock-coder. Ask about your code and the answer starts with the code. 986 MB on your laptop. No API key. Works with the Wi-Fi off. Guide: nockai.dev/guides/vscode

Video is where this hits hardest. It’s huge, it’s personal, and uploading it is the part people hate. If a model could answer questions about your videos without them leaving your laptop, what would you ask it first? We’re collecting use cases for the test set.

Mehdi @MehdiHacks

a week ago

Many people misunderstand local AI. They see it as a replacement for frontier models like Opus. That's missing the point. Local AI solves 3 main problems: 1. Privacy and data ownership: Full control over your data. Run LLMs on isolated or completely offline machines. 2. Control:

23 12 200 16K 73

We gave Nock Agent 31B, our fine-tune of @googlegemma’s Gemma 4 31B, a bug report and two tools. Before reading or editing anything, its first move was this: run_command: python -m pytest -q test_lists.py Reproduce first, then fix. That’s the habit we trained in.

Where to get Nock Agent 31B: LoRA for vLLM: huggingface.co/Nock-AI/nock-a… GGUF for llama.cpp (Q4_K_M, 19 GB): huggingface.co/Nock-AI/nock-a… Both open weights, Apache-2.0. Commands and results are on the model cards.

Runs on one 80 GB GPU with vLLM, on top of Gemma 4 31B. The serve command and the system prompt are in the model card. A 19 GB GGUF for llama.cpp is up too: huggingface.co/Nock-AI/nock-a…

How we test it: 50 small Python projects, each with an issue and a failing test. 20 of them get extra tests only after the agent is done, so a fix that just satisfies the visible test fails. 3 runs per task, 150 in total. Where it misses: tasks where the same fix is needed in several files. In 3 runs it renamed a function in one place and left two other modules on the old name. If your change touches many files, tell it to search the whole project first. Want it as one file? 19 GB GGUF for llama.cpp: llama-server -hf Nock-AI/nock-agent-31b-GGUF:Q4_K_M --jinja Given a bug report and two tools, its first move was to run the tests before touching the code.

Nock Agent 31B is out. A fine-tune of Gemma 4 31B for coding agent work: read the code, make the smallest fix, run the tests, stop. On our 150-run agent test it solved 145 and edited a test file 0 times. Open weights: huggingface.co/Nock-AI/nock-a…

The 30 ms code review is the part that stuck with me. A model can get the right answer and still skip checking its own work. We saw the same thing in our own agent tests: right fix, then “done” without running the tests again. Getting it right and checking your work are two separate skills.

Dan Shipper @danshipper

5 days ago

BREAKING: Gemini 4 is OUT and it's SURPRISINGLY GOOD?? I admit I was ready to trash this model qhen I saw the benchmarks but six of us at @every spent the past week testing it across coding, writing, design, and knowledge work, and were pleasantly surprsied. It’s a

71 27 424 99K 185