Ran Laguna S 2.1 FP8 on dual DGX Spark (TP=2, DFlash W4A16). @poolsideai @jasoncwarner Here are my first benchmarks, tips for running it and my review of this model. Benchmarks Couldn't recreate the 60.9 tok/s single-stream decode that's been floating around. (I think this is because num_speculative_tokens was set at k=6 not k=15) 😟 Single-stream (512 tok, code prompt, temp=0): - Mean: 33.1 tok/s (±1.1 across 5 runs) - Peak: 38.2 tok/s - No-draft baseline: 26.5 tok/s - DFlash speedup: 1.25x (recipe claims 2.3x) Concurrency sweep (256 tok, temp=0): - conc=1: 37.3 tok/s aggregate - conc=2: 62.4 tok/s aggregate - conc=4: 120.1 tok/s aggregate - conc=8: 150.0 tok/s aggregate, 0 failures - KV wall: not reached at conc=8 DFlash spec decode acceptance: 34.2% (recipe claims 62.2%). Tested 4 configs — W4A16 vs FP8 draft, draft_model vs dflash method, DeepGEMM vs triton backend, temp=0 vs 0.7. None moved acceptance significantly. onder how to get it to 62%" → something like "We ran k=6 (not the recipe's k=15) — that's likely part of the gap. But even at k=6, the community numbers on GB10 are consistent with ours. Flags and settings: vllm serve \ --model Laguna-S-2.1-FP8 \ --tensor-parallel-size 2 \ --distributed-executor-backend ray \ --gpu-memory-utilization 0.85 \ --max-model-len 262144 \ --max-num-seqs 8 \ --max-num-batched-tokens 8192 \ --kv-cache-memory-bytes 5368709120 \ --enforce-eager \ --enable-auto-tool-choice \ --tool-call-parser poolside_v1 \ --reasoning-parser poolside_v1 \ --speculative-config '{"model": "Laguna-S-2.1-DFlash-W4A16", "method": "draft_model", "num_speculative_tokens": 6}' ^ changing this to this to 16 may unlock more decode speed Key gotchas: - --max-num-seqs 8 is CRITICAL — without it DFlash OOMs the GPU under concurrency - --enforce-eager required — cross-node TP2 CUDA graph deadlocks on GB10 - --moe-backend triton is faster single-stream but crashes at conc=4 - Symlink model paths across nodes or Ray workers can't find the checkpoint - Patch vLLM's laguna_dflash.py for W4A16 draft (hasattr guard on get_cache_scale) - python3.12-dev required on both nodes for Triton JIT My Review: Why Laguna S 2.1 Crushes Qwen 27B Dense and DS4F on long horizon agentic tasks. I just put Laguna S 2.1 through its paces against Qwen 27B dense and DS4F, and honestly, it’s not even a close fight. For this specific workflow, Laguna absolutely outperforms them. Here is my breakdown of why it wins: 1. Smart Routing for Wildly Different Tasks My pipeline isn't just one repeatable task. It’s a chaotic mix of JSONL parsing, semantic search, browser DOM extraction, natural language synthesis, and style-constrained drafting. The Problem with Qwen 27B: Because it's a dense model, it uses every single parameter for every single token. It's verbally smart as a result, but the moment you start throwing long context tool chaining at it, the 27B starts reminding you that it is a 27B moddel at the end of the day. The Laguna Advantage: Even when compared to an elite sparse MoE like Qwen3.6-35B-A3B, Laguna S 2.1 holds a massive structural edge for deep pipeline work. Qwen's '35B total / 3B active' architecture makes it incredibly fast and dirt cheap to run, using linear attention tricks to zip through standard coding tasks. But Laguna scales the MoE trick to a whole different level with 118B total parameters and 8B active. When you are forcing a model through a brutal 25-step tool chain with 16 simultaneous logic constraints, Qwen’s aggressive 3B active bottleneck can start dropping fine details. Laguna's top-10 expert routing gives its token parsing just enough extra math 'heft' to keep the entire execution completely bulletproof across its 1-million token context window. 2. It Doesn't Crack Under Heavy Constraints This workflow throws 16 simultaneous constraints at the http://model. In my testing, the dense 27B model starts losing the plot and degrading around constraint 5 or 6. Laguna, on the other hand, holds all 16 constraints perfectly across a chain of 10+ tool calls. Its instruction-following experts are simply bigger and more robust. 3. Bulletproof Tool-Calling Reliability This was the real test: 25 sequential tool calls involving CDP navigation, Runtime.evaluate with JSON parsing, curl with the fxtwitter API, vault queries, and file reads.Laguna executed pretty close to flawless. Zero fumbled calls, zero hallucinated paths, and zero malformed parameters. Steering required 1 reminder that a tool was available. For agentic workflows, that is the only metric that actually matters. DS4F is great at coding , but its parameter construction for tool calls gets sloppy when you move outside of pure coding workflows. Laguna is just way more precise. Downside: Laguna is quite weak on voice adherence (subjectively, for pure chatbot intelligence Laguna isn't it) Initial Verdict: If you are building complex, multi-step agents that need to juggle tools and strict instructions, Laguna S 2.1 is the fantastic choice Recipe in reply 👇

For my first post, I’m sharing a letter @NVIDIA signed on why open models matter. AI will transform every industry, power every company, and be built by every country. Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. The world needs both frontier closed models and frontier open models. images.nvidia.com/pdf/Open-W…

“Tier 1” VC firms is the biggest psyop going in Silicon Valley. I see so many young founders falling for this shit. It’s marketing. The things to optimize for in a fundraise are control, dilution, and the individual person you will be working with at the firm, in that order.

I just learned that silent reading isn't silent for everyone. Some people hear the words in their heads as they read. 🤯

Billions in crypto stolen. If we can't stop the thefts, can we reduce the harmful effects? Over recent months, a couple other @Stanford researchers and I drew out and prototyped ERC-20R/721R to support reversible transactions on #Ethereum. See post & 🧵: mirror.xyz/kaili.eth/gB-rx89…

An open list of problems in DeFi, the problems we solve over the next 24-36 months decide our next level of adoption. mirror.xyz/0xemperor.eth/0gu…