
Founder, President & Chief Scientist @PrunaAI Prev. @Twitter research, Ph.D. in ML @TU_Muenchen
Joined January 2019
- Tweets787
- Following189
- Followers790
- Likes2K
Pinned Tweet
Today, we are happy to release P Video 2 Pro, Based on H3 from @MiniMax_AI .
Highest Quality, Highest Speed.
More benchmarks (and more models?!) coming soon!
Today, we launch P-Video-2-Pro, based on MiniMax H3: The best quality. The best speed.
Text or image in. Video and audio out.
โข ๐๐ป๐ฝ๐๐: Text or first-frame image, with optional last-frame conditioning
โข ๐ฆ๐ฝ๐ฒ๐ฒ๐ฑ: ~2.0s for a 5s clip at 480p ยท ~4.3s at 768p in Speed mode
โข ๐ฃ๐ฟ๐ถ๐ฐ๐ถ๐ป๐ด: From $0.02/s Speed ยท $0.04/s Quality
โข ๐ ๐ผ๐ฑ๐ฒ๐: Speed or Quality
โข ๐๐ผ๐ป๐๐ฟ๐ผ๐น: First and last frame, seed, aspect ratio, and prompt upsampling with Off, Turbo, or Max
Available via our inference partners @replicate, @Cloudflare, @eachlabs, @scenario_gg, @TellersAI, @wavespeed_ai, @wiroai, @runware, and @togethercompute
Validated by our benchmark partners @datapointai, @RapidataAI, and @DesignArena. Benchmarks to be released soon ๐
๐ Free Playground: playground.pruna.ai/
๐ Model docs: docs.api.pruna.ai/guides/modโฆ
๐งฉ API: dashboard.pruna.ai/login
Bertrand Charpentier retweeted
Academia has a problem:
Every author wants to use LLMs to write papers, but no reviewer wants to read AI (assisted) content.
And every reviewer wants to use LLMs for reviews, but no author wants their paper judged by an AI.
Authors & reviewers are the same people though...
Bertrand Charpentier retweeted
anthropic.com/research/invesโฆ
just read this and i think it is very inspiring. we care too much about the so called capabilities but care too less about the unintended actions that might cause big troubles. if anyone thinks about new evals, you can build towards this direction
Bertrand Charpentier retweeted
What if the future was green, quiet, and endlessly vast? Solarpunk concept film ๐ฌ
made with P-Video-2-Pro-Reference by @PrunaAI
P-Video-Upscale just got faster and cheaper! โก
Introducing Turbo mode: faster video upscaling at 65% lower cost compared to Standard (up to 30 fps), with a slight trade-off in quality.
Here's what you need to know:
- ๐ฃ๐ฟ๐ถ๐ฐ๐ถ๐ป๐ด: Starting at just $0.0035/s
- ๐ฆ๐ฝ๐ฒ๐ฒ๐ฑ: Faster processing across resolutions
- ๐ฅ๐ฒ๐๐ผ๐น๐๐๐ถ๐ผ๐ป: 720p, 1080p, 2K, or 4K
- ๐๐ฟ๐ฎ๐บ๐ฒ ๐ฟ๐ฎ๐๐ฒ: Up to 60 fps, at no extra cost in Turbo mode
Available via our inference partners @scenario_gg, @replicate, @WiroAI, @TellersAI, @eachlabs, @wavespeed_ai, Sociaro, and @runware.
๐ Model docs: docs.api.pruna.ai/guides/modโฆ
๐งฉ Try it through the API or Playground: dashboard.pruna.ai/login
Bertrand Charpentier retweeted
Pruna's color is purple, and so is my heart. ๐
So this Christmas, I'm skipping red and green and dressing it all in lavender.
Images made with P-Image-Ideogram on @PrunaAI
Cheaper agents no longer just need a smaller model: what they read, the harness around them, how they team up, where they run, and how little post-training has to change.
๐๏ธ Weekly efficiency digest: 5 papers on agent efficiency beyond the model. ๐งต
1
5
215
4๏ธโฃ Agents can build specialized on-device models: Given a task specification, Mobile AI Factory fine-tunes + quantizes an SLM per task and phone. On unseen tests it beats 5-shot Claude Sonnet 5 & DeepSeek v4 Flash on 5/10 tasks, ties on 3. E
arxiv.org/abs/2611.13509
1
36
5๏ธโฃ Pick of the week โญ Test-time compute scales better with parallel agent teams. Parallel teams sample small agent teams and keep the best team result. 10 teams of 10 beat (1) best-of-100 agents on 10/12 tasks (+27% avg) and (2) one 100-agent team on 4/4 (+8โ30%).
arxiv.org/abs/2611.09798
2
14
How does P-Video-2-Pro-Reference perform across quality, speed, and cost? The latest P-Bench results show it sitting on all three Pareto frontiers.
Across 7 reference-to-video models, the Ref2V Bench v0 results give:
โข Cost: #1 ยท ~$0.046 per output video second (including reference fees)
โข Quality: #2 ยท 1033 Elo (Rapidata)
โข Speed: #2 ยท 2.28s per output video second
P-Video-2-Pro-Reference delivers the lowest cost in the benchmark, while ranking second in both human preference and generation speed.
It sits on the preference ร speed, preference ร price, and speed ร price Pareto frontiers.
๐ฅ Full benchmark: huggingface.co/spaces/PrunaAโฆ
I have been fine tuning the H3 Image decoder overnight. I added a facial identity and OCR loss on top of lpips, fft, and l1 losses. I am not doing a critic loss. P-H3-Image VAE still has better metrics, but it is also a significantly larger decoder. Keeping it going.
We open LoRAs to precisely remove objects (and their reflections) in seconds.
It works with our previous distillation of Qwen 2.1 from @Alibaba_Qwen.
Try it here
- huggingface.co/PrunaAI/P-Q-Iโฆ
- huggingface.co/PrunaAI/P-Q-Iโฆ
Today, we open-source P-Q-Image-2.1-Remove-Object, 3 LoRA adapters for Qwen-Image-2.1 by @Alibaba_Qwen. Remove the object you point at, along with its shadow or reflection, and change nothing else!
- ๐๐ผ๐ ๐ฒ๐ณ๐ณ๐ถ๐ฐ๐ถ๐ฒ๐ป๐ ๐ถ๐ ๐ถ๐? Remove objects in 6 steps instead of 40, with no CFG. Use the 6-step adapter as the default. An 8-step adapter is also available, and a 40-step adapter runs only on the base model, without the few-step model's drift.
- ๐๐ผ๐ ๐ด๐ผ๐ผ๐ฑ ๐ถ๐ ๐ถ๐? Shadows and reflections just go away along with the object, even when they sit outside the outline. The rest of the photo stays put and lines up with the input, so you can use the result as is.
- ๐๐ผ๐ ๐ฑ๐ผ๐ฒ๐ ๐ถ๐ ๐๐ผ๐ฟ๐ธ? The adapters load directly on top of Qwen-Image-2.1, so the pipeline stays unchanged. The few-step variants stack on our P-Q-Image-2.1 distillation LoRA. Mark the object with a rough red outline, or just name it in the instruction.
๐ Blog: pruna.ai/blog/p-q-image-2.1-โฆ
๐ค Open Weights: huggingface.co/PrunaAI/P-Q-Iโฆ
โก๏ธ ComfyUI: huggingface.co/PrunaAI/P-Q-Iโฆ
Bertrand Charpentier retweeted
Riding the elevator with young people whose style is too advanced ๐ณ
Grandpa just needed one word: "Wow."
Let the elevator do the rest. ๐โจ
made with P-Video-2-Pro-Reference by @PrunaAI
Prompt ๐
The most interesting Muse Ember number isn't the price.
- Ember (max): 64% declined ยท 21.1% accurate ยท 17.4% hallucination
- Haiku 5.5 (max): 40% declined ยท 34.7% accurate ยท 34.2% hallucination
- Gemini 3.8 Flash (high): 19% declined ยท 50.5% accurate ยท 55.1% hallucination
Saying "I don't know" when unsure can lower hallucinations (wrong / non-correct) and caps accuracy (correct / all).
Meta's newest model Muse Ember scores 36 on the Artificial Analysis Intelligence Index, matching Claude Haiku 5.5 (max) at a third of the Cost per Task
Muse Ember is Meta's newest efficient-tier model line, with the same input and output token prices as Claude Haiku 5.5 and GPT-6 Luna
Key Takeaways:
โคย On the Intelligence Index vs. Cost per Task Pareto frontier at both max and xhigh effort.ย Muse Ember (max) is the cheapest model scoring 36 or higher, at $0.14 per task, and Muse Ember (xhigh) scores 34 for $0.12. Claude Haiku 5.5 (max) costs almost 3x more at $0.41 for the same score as Muse Ember (max). Muse Ember (max) is also about 4x cheaper than DeepSeek V4.1 Flash 1017 (max, 33, $0.60) and GLM-5.3-Flash (28, $0.58)
โคย 8 points behind Muse Wonder for about 1/50th of the Cost per Task.ย Muse Wonder (max) scores 44 at $6.95 per Intelligence Index task
โคย Stands out on AutomationBench-AA.ย Muse Ember (max) scores 68.1% on AutomationBench-AA, level with Muse Wonder (max), and beats Claude Haiku 5.5 (max) on GDP.pdf (22.8% vs 20.8%), but trails it on AA-Briefcase (1521 vs 1574 Elo)
โคย Mixed in agentic coding in our Intelligence Index evaluations.ย It beats Claude Haiku 5.5 (max) on Terminal-Bench 4.0 (40.9% vs 32.8%) but trails it on FrontierCode 2.0 (27.2% vs 34.9%)
โคย Ahead of similarly priced models when running in its own coding agent.ย In Muse Code, Muse Ember (max) scores 46 on the Artificial Analysis Coding Agent Index, ahead of Claude Haiku 5.5 in Claude Code (max, 37) and GPT-6 Luna in Codex (max, 31)
โค Knowledge is the main weakness, but it rarely hallucinates. Its AA-Omniscience accuracy of 21.1% is the lowest of the similarly intelligent models we compared, against 34.7% for Claude Haiku 5.5 (max) and 50.5% for Gemini 3.8 Flash (high). This is mostly because it declines to answer 64% of questions, which keeps its hallucination rate at 17.4%, half of Claude Haiku 5.5โs (max)
Other details:
โคย Pricing:ย $0.10 input / $0.004 cached input / $0.50 output per 1M tokens
โคย Reasoning effort:ย results shown at max effort, with xhigh also on the Intelligence vs. Cost chart โคย Availability:ย Meta's first-party API and Muse Code
1
2
134
Bertrand Charpentier retweeted
There's a whole universe wrapped in twilight inside this wooden glass case.
A rare lavender moth rests under a soft glowing nebula, breathing in stardust.
Created with P-Image by @PrunaAI
Pruna v0.3.5 is out! ๐
Here are a few highlights:
- New FP8 quantization algorithms: static FP8, and time-aware FP8 for diffusion pipelines.
- New MSE metric in the evaluation toolbox
- Tighter security: safer checkpoint loading and a 7-day dependency cooldown
- New tutorials for FP8 quantization and sustainability benchmarking
๐ Read the notes: buff.ly/GnDidPi
Thank you all for your contribution!
New challenge to celebrate the P-Video-2-Pro Series! $2,000 Prize Pool!
๐ญ ๐๐๐๐ ๐๐๐ ๐๐๐๐. ๐๐๐๐๐ ๐๐๐ ๐๐๐๐๐.
๐ฃ-๐ฉ๐ถ๐ฑ๐ฒ๐ผ-๐ฎ-๐ฃ๐ฟ๐ผ ๐๐ฒ๐ฟ๐ถ๐ฒ๐. Generate the scene, lock the character, deliver the line, or recast the actor. Pick the model that fits the job.
Cut a vertical microdrama hook that proves quality, speed, or price.
Beyond compression algorithms, compressing a model is also a system problem: hardware budgets, serving stacks, energy, and the order in which you stack methods.
๐๏ธ Weekly efficiency digest: 5 papers on measuring and optimizing the whole system. ๐งต
4๏ธโฃ Compression Technique Interactions in SLMs: After testing 72 compression configurations, the study shows that compression methods do not combine independently. E.g. quantization + pruning can lead to stronger collapse. Fine-tuning is important for recovery.
arxiv.org/abs
1
24
5๏ธโฃ Pick of the week โญ CompressionBench. It formulates LLM compression as a terminal task where coding agents build weights + inference code under storage limits. Optimization agents were capable of maintaining 77.33% mean accuracy vs 79.56% for BF16.
arxiv.org/abs
1
23
















