
Founder, President & Chief Scientist @PrunaAI Prev. @Twitter research, Ph.D. in ML @TU_Muenchen
Joined January 2019
- Tweets789
- Following189
- Followers791
- Likes2K
5️⃣ Pick of the week ⭐ 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 (+8–30%).
arxiv.org/abs/2611.09798
2
17
4️⃣ 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
41
3️⃣ Fine-tuning works even if we simply delete a few weights: Deletion Tuning learns a binary mask zeroing <0.05% of weights. On Qwen3-30B-A3B with GRPO: 58.54% for deletion fine-tuning vs 58.43% for full fine-tuning, with a 29.6 MB mask vs 116,471 MB checkpoint (~3,900× smaller).
arxiv.org/abs/2611.10866
1
18
2️⃣ Cost-Efficient Agent Harnesses: the harness barely moves accuracy but moves cost up to 2.6× on the same model and task. Re-reading context is up to 73% of cost. Their agent layer on Claude Code/Codex cuts cost up to 35% with marginal accuracy loss.
arxiv.org/abs/2611.12122
1
34
1️⃣ Agents don't need to read full documents; they can read sparsely. SparseRead shows them only the evidence they need. Sparse reading vs full reading in 30 settings: median 61.4% fewer tokens, 53.7% less wall time (up to 92.9% / 89.0%), score kept or better in 29/30.
arxiv.org/abs/2608.22237
1
26
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-I…
- huggingface.co/PrunaAI/P-Q-I…
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!
- 𝗛𝗼𝘄 𝗲𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝘁 𝗶𝘀 𝗶𝘁? 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.
- 𝗛𝗼𝘄 𝗴𝗼𝗼𝗱 𝗶𝘀 𝗶𝘁? 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.
- 𝗛𝗼𝘄 𝗱𝗼𝗲𝘀 𝗶𝘁 𝘄𝗼𝗿𝗸? 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.
👉 Blog: pruna.ai/blog/p-q-image-2.1-…
🤗 Open Weights: huggingface.co/PrunaAI/P-Q-I…
⚡️ ComfyUI: huggingface.co/PrunaAI/P-Q-I…
The most interesting Muse Ember number isn't the price.
- Ember (max): 64% declined · 21.1% accurate · 17.4% hallucination
- Haiku 5.5 (max): 40% declined · 34.7% accurate · 34.2% hallucination
- Gemini 3.8 Flash (high): 19% declined · 50.5% accurate · 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:
➤ On the Intelligence Index vs. Cost per Task Pareto frontier at both max and xhigh effort. 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)
➤ 8 points behind Muse Wonder for about 1/50th of the Cost per Task. Muse Wonder (max) scores 44 at $6.95 per Intelligence Index task
➤ Stands out on AutomationBench-AA. 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)
➤ Mixed in agentic coding in our Intelligence Index evaluations. 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%)
➤ Ahead of similarly priced models when running in its own coding agent. 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)
➤ 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 5.5’s (max)
Other details:
➤ Pricing: $0.10 input / $0.004 cached input / $0.50 output per 1M tokens
➤ Reasoning effort: results shown at max effort, with xhigh also on the Intelligence vs. Cost chart ➤ Availability: Meta's first-party API and Muse Code
1
2
136
5️⃣ Pick of the week ⭐ 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
4️⃣ 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
3️⃣ The SIMPLE Law predicts inference energy from architecture and sequence lengths. Four fitted parameters, 13 dense LLMs on H100: 13.25% mean absolute percentage error when holding out an LLM (i.e. no per-model fit).
arxiv.org/abs
1
11
2️⃣ ServerBench: a benchmark for designing complete LLM serving systems to maximize service value per GPU dollar. E.g., across 11 coding-agent models and
three deployments, Claude Sonnet 5.5 achieves 2.97 dollars per dollar on Gemma E4B, 4.87× the vLLM serving reference.
arxiv.org/abs
1
32












