Reasoning models can now drive robots. But the robot sits still while the model thinks.
SpecPilot now speeds up agentic robotics. On PEAR (Qwen3.8-27B as a robot policy), the same LIBERO task finishes in 49 s instead of 189 s.
3.84× faster task completion
71 → 241 tok/s decode
No training 🤖
RSI starts with infra 🔁 mlsys is in a golden age. agents are great researchers now. with tokens, "low-hanging fruit" is everywhere.
agents optimize speculative decoding for parallel drafters (DFlash2, DSpark) without training or quality loss. their 12-bit format beats cuBLAS.
Reasoning models can now drive robots. But the robot sits still while the model thinks.
SpecPilot now speeds up agentic robotics. On PEAR (Qwen3.8-27B as a robot policy), the same LIBERO task finishes in 49 s instead of 189 s.
3.84× faster task completion
71 → 241 tok/s decode
No training 🤖
RSI starts with infra 🔁 mlsys is in a golden age. agents are great researchers now. with tokens, "low-hanging fruit" is everywhere.
agents optimize speculative decoding for parallel drafters (DFlash2, DSpark) without training or quality loss. their 12-bit format beats cuBLAS.
Exciting to see what DFlash2 and SpecPilot can achieve together. Parallel drafting and systems optimization go hand in hand, with so much more potential to unlock. We believe parallel drafters are the future.
Reasoning models can now drive robots. But the robot sits still while the model thinks.
SpecPilot now speeds up agentic robotics. On PEAR (Qwen3.8-27B as a robot policy), the same LIBERO task finishes in 49 s instead of 189 s.
3.84× faster task completion
71 → 241 tok/s decode
No training 🤖
The reasoning rule alone makes the robot clock 1.87× faster, with the same success rate (81.7% vs 81.7%) over 120 paired episodes.
The rule is lossy and validated on LIBERO only. Code and reproduction scripts are in the repo.
🌐 https://t.co/HuFwg0R6ER
💻 https://t.co/SpXTHw7XFB
How: PEAR runs on the SpecPilot engine (W12 weights, tree verification, two-level verification), plus an efficient reasoning rule:
• accept drafted tokens the target gives p ≥ 0.3
• skip hesitation words ("Wait", "Hmm")
• never relax inside a repeated sequence
Median speedup over the PEAR baseline: 4.04× across 19 LIBERO episodes.
The reasoning rule alone makes the robot clock 1.87× faster, with the same success rate (81.7% vs 81.7%) over 120 paired episodes.
The rule is lossy and validated on LIBERO only. Code and reproduction scripts are in the repo.
🌐 https://t.co/HuFwg0R6ER
💻 https://t.co/SpXTHw7XFB
How: PEAR runs on the SpecPilot engine (W12 weights, tree verification, two-level verification), plus an efficient reasoning rule:
• accept drafted tokens the target gives p ≥ 0.3
• skip hesitation words ("Wait", "Hmm")
• never relax inside a repeated sequence
Median speedup over the PEAR baseline: 4.04× across 19 LIBERO episodes.
We breathed life into LLM system.
Introducing Evolving Infra⚙️: an inference framework that adapts diverse and dynamic agent workload purely online and retires fixed PD disaggregation.
vs. NVIDIA Dynamo on up to 256 GB300 GPUs:
➡️ Up to 4.1× lower latency
➡️ Up to 14× lower P95 TTFT
➡️Up to 2.36× more revenue
on real agent traces(Computer Use, Coding and math agents) and opensource models.
🌐 Website: https://t.co/7I3iERjE6b
📄 Tech report: https://t.co/qDGcwl0svx
💻 Code: https://t.co/0VOBCptKcb
Which agent(e.g. Pi/Openhands) do you want us to accelerate next? Tell us below 👇
Yes, 10.87× is with one request. With batching, the stock baseline's speedup over AR roughly halves, from 4.46× to 1.94×, but SpecPilot still reaches 3.93×, about 2× over stock. Once the batch is large, every verified row costs real compute, so the loop optimizes for that regime separately. It spends rows only where acceptance is likely and removes costs that don't scale with rows, such as separate prefills that stream the weights again.
SpecPilot is out 🚀
we gave coding agents GPUs and one rule: stay lossless. with one request, they more than doubled every released speculative decoding system we tested.
DFlash2: 4.46× → 10.87× over AR. no training. plus a lossless 12-bit format whose GEMM beats cuBLAS.
RSI starts with infra 🔁 mlsys is in a golden age. agents are great researchers now. with tokens, "low-hanging fruit" is everywhere.
agents optimize speculative decoding for parallel drafters (DFlash2, DSpark) without training or quality loss. their 12-bit format beats cuBLAS.
Thanks! Both, by design. Speculative decoding is deployment-specific, so the loop re-optimizes per deployment (regime, drafter, target, GPU). What transfers is a structured memory of validated mechanisms. DSpark reused the tree, W12 and TLV entries, and EAGLE-3 on Gemma-4-12B, a sequential drafter, reimplemented W12 and prompt admission from memory. Each reuse is re-tuned and re-timed.
On W12: it's bit-exact, not approximate. Most BF16 weights in a channel fit a 16-binade window, so sign + 4-bit exponent offset + 7 mantissa bits = 12 bits, and outliers go to an exact escape list. Every tensor is checked bit for bit at load. The win is in the memory-bound regime: 25% fewer bytes, decoded in registers. On H100 the GEMM beats cuBLAS by 1.19–1.47× at 1 row and 1.17–1.35× at 32 rows on Qwen3.8-27B projections.
Good question! The format itself is ~12.01 bits/weight, including escapes and per-channel metadata, so the converted layers do shrink ~25% (Qwen3.8-27B: 44.67 → 33.58 GB). The gap is what we don't convert: W12 covers the decoder linear layers and LM head, everything else stays BF16. Escapes are <0.013% of weights.
RSI starts with infra 🔁 mlsys is in a golden age. agents are great researchers now. with tokens, "low-hanging fruit" is everywhere.
agents optimize speculative decoding for parallel drafters (DFlash2, DSpark) without training or quality loss. their 12-bit format beats cuBLAS.