Laguna S 2.1: 8B Active Params, 70% Terminal-Bench

Poolside released Laguna S 2.1 on July 21, and the number that matters is not the benchmark score, it is the activation count. This is a 118-billion-parameter mixture-of-experts model that lights up roughly 8 billion parameters per token, posts 70.2% on Terminal-Bench 2.1, ships open weights on Hugging Face, and sells for $0.10 per million input tokens on OpenRouter. That combination, frontier-adjacent agentic coding at small-model economics with weights you can download, is the interesting claim. It is also, for now, a claim: Poolside published no SWE-bench Verified figure, so Laguna does not enter our ranked leaderboard.

• 118B total, ~8B activated per token, with a 1M-token context window. Inference cost tracks the activated parameters, not the total.
• Terminal-Bench 2.1: 70.2% with thinking enabled, up from 60.4% without. SWE-bench Pro: 59.4%. SWE-bench Multilingual: 78.5%.
• Open weights under the OpenMDW-1.1 license in BF16, FP8, INT4 and NVFP4, with GGUF and MLX conversions and vLLM, SGLang and Ollama support.
• Training began May 22, 2026 and shipped in under nine weeks, reusing XS 2.1’s pre-training data. The gains are post-training and context, not scale.

What is actually new in 2.1?

Poolside is explicit that S 2.1 reuses the same pre-training data as XS 2.1. No new corpus, no scale jump. The improvements come from post-training and a longer context window, and the way the team describes the target behaviour is unusually concrete for a model card: more verification, less taking things for granted, not declaring victory early, and being more persistent.

RelatedKimi K3 Is the Largest Open Model Ever, and It Is Not Cheap

Those are agent failure modes, not capability gaps, and the benchmark deltas back that reading. Terminal-Bench 2.1 moves from 60.4% to 70.2% when thinking is enabled. DeepSWE v1.1 moves from 16.5% to 40.4%, which is a much larger swing and tells you the base capability was largely there while the harness behaviour was not. A model that stops early, assumes a command worked, or declares a task done without checking will fail long-horizon agentic benchmarks regardless of how much it knows. Fixing that is a post-training problem, and it is where a lot of the practical gap between models currently lives.

The nine-week turnaround from training start on May 22 to release on July 21 is worth noting on its own. That cadence only works if pre-training is amortised across a family and each release is a post-training pass, which is precisely the structure Poolside has been running from Laguna M.1 in April through XS 2.1 earlier this month.

Why does the activation count matter more than the score?

In a mixture-of-experts model the router selects a small subset of experts for each token, so the compute you pay for scales with activated parameters while memory scales with the total. Laguna S 2.1 holds 118B and activates about 8B, roughly a fifteen-to-one ratio, and that ratio is the entire economic story.

It shows up directly in the price. On OpenRouter the model runs $0.10 per million input tokens and $0.20 per million output, with Baseten and the Vercel AI Gateway also serving it. Against frontier coding models that bill in single or double dollars per million output tokens, this is one to two orders of magnitude cheaper for a model claiming 70.2% on Terminal-Bench 2.1. For agentic workloads, where a single task can burn hundreds of thousands of tokens across dozens of turns, that difference decides what is economically viable rather than merely which model is best.

The open weights compound it. The release is on Hugging Face under the OpenMDW-1.1 license with BF16, FP8, INT4 and NVFP4 checkpoints, plus GGUF and MLX conversions and support in vLLM, SGLang and Ollama. An 8B-activated model in INT4 is genuinely local-hardware territory, which puts a plausible agentic coding model inside environments that cannot send code to an API at all. You can also try it free at chat.poolside.ai before committing to anything.

Why is it not on our leaderboard?

Because Poolside did not publish a SWE-bench Verified score, and we do not rank models on scores nobody independent has reproduced.

Our AI coding leaderboard ranks on SWE-bench Verified and requires a source we can point at, preferring independent evaluation from vals.ai, which runs every model itself on one neutral harness. Laguna S 2.1 reports Terminal-Bench 2.1, SWE-bench Pro, SWE-bench Multilingual, DeepSWE, SWE Atlas and Toolathlon, all of them useful and none of them the axis the board ranks on. So it enters the verifying queue with a null score until an exact-match independent evaluation exists, and it will be re-checked on our daily sweep.

RelatedMurati's Thinking Machines Ships Inkling Open-Weights Model

This is not pedantry about a missing row. Vendor-reported SWE-bench figures have run between 2.6 and 11.6 points above neutral-harness measurements in the cases we track, because these benchmarks score a model together with its scaffolding, and vendors evaluate with their own. Poolside’s published numbers are self-reported by the same convention every lab uses; that makes them a reasonable starting estimate and not a verified result. Anyone quoting 70.2% as an established fact today is quoting a vendor.

What it means for the market

The signal here is margin compression at the cheap end of coding inference, not a threat to any frontier lab’s flagship. If an 8B-activated open-weights model genuinely delivers near-70% Terminal-Bench performance, the commodity tier of agentic coding, where volume is high and price sensitivity is extreme, gets very hard to charge premium rates for. Watch inference providers and API resellers whose pricing assumes a floor that open weights at $0.10 per million erases, and watch whether the frontier labs respond with cheaper small models rather than price cuts on flagships.

Our take

The most credible thing about this release is what Poolside chose to optimise. Reusing XS 2.1’s pre-training data and spending the effort on verification behaviour, persistence, and not declaring victory early is a bet that the binding constraint on agentic coding is discipline rather than knowledge. The DeepSWE jump from 16.5% to 40.4% with thinking enabled is the strongest evidence that the bet is right, and it is a more interesting result than the Terminal-Bench headline.

The economics deserve the attention they will get. Open weights, 8B activated, INT4 checkpoints and $0.10 per million input tokens is a genuinely different offer from the frontier tier, and for high-volume agent work the relevant comparison was never best available model, it was best model you can afford to run in a loop. That said, we are not putting a number on the board that one party measured. The absent SWE-bench Verified score is the gap between a very promising release and a verified one, and the honest position today is that Laguna S 2.1 looks excellent and remains unconfirmed. We will re-check it daily and rank it the moment somebody neutral runs it.

Original analysis by GenZTech. Benchmark figures as published by Poolside and vendor-reported unless independently confirmed. Current as of July 2026.