Alex

head of marketing at @atomic_chat_hq

[Alex @AlexFromAtomic]

I tested gemma 12b on my 16gb MacAir and it's very hot! Takes about 9GB in RAM btw

[atomic.chat @atomic_chat_hq]

New Google Gemma 4 12B claims near-26B performance - we tested both!

We ran both models locally on one RTX 4090 and gave each the same task: write a self-contained HTML5 canvas animation with real physics in one file without libraries. Three scenes - a Galton board, two blocks colliding off a wall, and a chaotic triple pendulum

Outputs:
Gemma 4 26B-A4B: 15 GB VRAM usage, 6.9k tokens, 138 tok/s
Gemma 4 12B: 9 GB VRAM usage, 8.9k tokens, 80 tok/s

Same Gemma 4 family, but the 26B-A4B won every scene and ran ~1.7x faster - on just 4B active params. The 12B stayed very close though, on almost half the VRAM - which makes it the ideal model for a 16 GB laptop

[Alex @AlexFromAtomic]

I’m shocked.. Nemotron has same performance but 10x cheaper? @NVIDIAAI cooked smth rly HOT!

[atomic.chat @atomic_chat_hq]

Nemotron 3 Ultra performed GPT 5.5 level 10× cheaper

We gave three same prompts to build HTML5 canvas with real physics. At first scene we have water in a spinning drum. Galton board - balls through pegs into bins. And a block collision setup with extreme mass differences.

Outputs:
Nemotron 3 Ultra: 11.3k tokens, $0.051
GPT 5.5: 11.0k tokens, $0.57

Nemotron stays right on GPT 5.5's heels, but at 10× cheaper. The gap in quality is far smaller than the gap in price.

[Overchat AI @OverchatAI]

I made @OpenAI GPT-5.5 and @GoogleDeepMind Gemini 3.1 Pro play REAL UNO — 10 games, full rules.

Gemini overthought every single card — literally 6× more reasoning tokens (133.5K vs 22.5K).

And it paid off — Gemini won 8–2.

Best part: they roast each other between cards the whole match 😂

Full stats ↓

[Alex @AlexFromAtomic]

@bayard0 @atomic_chat_hq the chaotic triple pendulum? ye that scene goes wild - real verlet, both models had to keep it stable through the chaos

[Alex @AlexFromAtomic]

@thealokverse @atomic_chat_hq @EXM7777 the 26B-A4B at 4B active is the fun part - 1.7x faster than the dense 12B and still won the physics scenes

[Alex @AlexFromAtomic]

@Ariadnavozz @atomic_chat_hq for tool/mcp calls gemma's genuinely decent, the 26B-A4B especially. basic coding it handles fine. frontier-level coding it's not - that's not what these sizes are for

[Alex @AlexFromAtomic]

@10xerik @atomic_chat_hq close on the easy scenes. the splits show on the hard physics - 26B led on speed and the pendulum, the 12B handled the block collision better. not a clean sweep either way

[Alex @AlexFromAtomic]

@aiseomastery @atomic_chat_hq exactly the point - ~9 GB, fits a 16GB laptop with room to spare. that's what google built the 12B for

[Alex @AlexFromAtomic]

@prodbitz @atomic_chat_hq @atomicbot_ai ye thats the read - 4B active is why it ran ~1.7x faster than the dense 12B and still took the scenes. MoE doing its thing

[Alex @AlexFromAtomic]

@Rom609033637850 @atomic_chat_hq we didn't run qwen here, it was gemma vs gemma. qwen comparison is the most requested though, doing it next. which qwen would you put up against it?

[Alex @AlexFromAtomic]

@AlanAiEngineer @atomic_chat_hq qwen's likely ahead on raw quality, fair. one note though - neither ran in 3 GB here, the 12B was ~9 GB and the 26B-A4B ~15 GB on a 4090. which gemma/quant did you try at 3?

[Alex @AlexFromAtomic]

@justjayvi @atomic_chat_hq qwen's next, easily the most requested. what size you running locally?

[Alex @AlexFromAtomic]

@NikiBelokopytov @atomic_chat_hq could well be. memory-matched is the fair follow-up, we'll cut the 26B to the same budget and rerun. if the hobbled one still wins that's a real result

[Alex @AlexFromAtomic]

synced up metronomes is so satisfying

[atomic.chat @atomic_chat_hq]

StepFun Step 3.7 Flash smashed DeepSeek V4-Flash in a physics contest

We gave two open-weight models the same task: write a self-contained HTML5 canvas animation with real physics in one file without libraries. Three scenes - a Galton board, balls bouncing in a spinning hexagon and five metronomes that sync up

Outputs:
Step 3.7 Flash: 59.6k tokens, 9m 57s
DeepSeek V4-Flash: 52.5k tokens, 6m 21s

DeepSeek was faster, but that's all it had. StepFun's model won on every front: physics simulation, visuals, and logical rendering of each scene

[Andy D @hypersonq]

I ran a 10-game AI chess tournament: @claudeai Opus 4.8 vs @MiniMax_AI M3

API calls powered by @aimlapi

In this game, Claude crushes MiniMax in 28 moves
Claude costs ~8x more per move: ~$0.020 vs ~$0.0025
Total: ~$0.62

Full results below ↓

[Alex @AlexFromAtomic]

@JohnShina91 @atomic_chat_hq different tests keep surprising us. what's the wildest one you've seen?

[Alex @AlexFromAtomic]

@adityawaslost @atomic_chat_hq not AGI yet 😄

[Alex @AlexFromAtomic]

@Umar__786Ai @atomic_chat_hq get your 5090 ready ;) open weights next week

[Alex @AlexFromAtomic]

@YLyGZn4hLxLXZXS @atomic_chat_hq sure it knows the genre, but knowing it and reading a messy napkin into working code are two different things i guess

[Alex @AlexFromAtomic]

@RamanduLight @atomic_chat_hq @doodlejump haha it's Doodle Jump style, not a copy - big difference :))