Fast and affordable inference for coding and research.
United States
Joined October 2026
- Tweets32
- Following12
- Followers18
- Likes9
New on motionbench: a contributed task from @ConorBronsdon. A narrated statistic has to land on its spoken cue, then keep its qualifier and source readable.
Every model hit the cue; the misses were all layout. It stays out of the score for now.
motionbench.sh
CoralBricks Community retweeted
My code review loop: use multiple model families to challenge shared blind spots:
- Cheap checks first (cut costs)
- Independent cross-family review (catch biases)
- Reproduce findings
- Re-review final commit post fixes
- Evidence drives merge decisions
Agents code. Humans steer.
CoralBricks Community retweeted
Really proud of the team for getting motionbench out: seven AI models, real motion-graphics briefs, every frame measured. Opus 5.5 leads, with Sol and DeepSeek close behind.
It's early, and we want it to mature in the open. Tell us what it gets wrong.
xitter.cf/coralbricksdev/status/…
ALT motionbench: score against cost per task for seven AI models. Claude Opus 5.5 leads at 0.987; GPT-6.1 Sol 0.981 and DeepSeek V4.1 Flash 0.968 are within its margin.
Can a model animate?
We built motionbench: 7 AI models write motion graphics as code, and headless Chrome measures every frame. Squash and stretch, arcs, anticipation, beats, camera pans. No model judges.
Results, live players, open data: motionbench.sh
Can a model animate?
We built motionbench: 7 AI models write motion graphics as code, and headless Chrome measures every frame. Squash and stretch, arcs, anticipation, beats, camera pans. No model judges.
Results, live players, open data: motionbench.sh
All open: the tasks, the grader, every composition each model wrote, and every frame-by-frame check.
Code: github.com/Coral-Bricks-AI/m…
Data: huggingface.co/datasets/cora…
Get a motion graphics film from any X post within seconds.
The model doesn't render video. It writes the film as HTML and JSON, one scene per line, and a small helper to animate them the moment it lands. The whole film finishes at roughly the same time a page loads.
One film per post (so we don't fry our GPUs) and we added rarity because it looked fun.
Try it on any post: coralreel.up.railway.app/new
GLM 5.3 support ends tomorrow
We’re retiring GLM 5.3 and moving that compute capacity to DeepSeek V4.1 Flash, now updated to the latest checkpoint.
If you’re still using GLM, please switch your model setting to deepseek-v4.1-flash-fast and test it with your usual workload before support ends. Your existing API key will work.
coralbricks.ai/models/deepse…
We know this is short notice. Let us know if you have any concerns
Now hosting DeepSeek-V4.1-Flash-1017.
Decode speed up to ~1,100 tok/s (p95, single stream) on SWE-bench Verified agentic coding replays, ~11K-token contexts.
The release ships a Python prompt encoder instead of a Jinja chat template.
Switching AI shouldn’t be this much of a stunt. 💀
Coral Bricks makes it easy.
coralbricks.ai
#AI #Inference
$20 of free credit on every new Coral Bricks account, while GPU capacity lasts. 🪸
Enough to put Claude Code or OpenCode on GLM 5.3 or DeepSeek V4.1 Flash for real work, not a demo.
coralbricks.ai/?utm_source=x…
ALT $5 crossed out, an arrow, then $20 in coral: free trial credit for every new account. Coral Bricks, coralbricks.ai.
Free trial just went from $5 to $20. 🪸
Every new Coral Bricks account starts with $20 of credit: millions of tokens of DeepSeek V4.1 Flash or GLM 5.3 for your coding agent.
Grab a key. Set the base URL. Build.
coralbricks.ai/?utm_source=x…
ALT $5 crossed out, an arrow, then $20 in coral: free trial credit for every new account. Coral Bricks, coralbricks.ai.
Heads-up: the $20 runs while we have spare GPU capacity, so it can end without notice. Grab it while it's on.