Fleet Watch Brief · 2026-07-22

Laguna S 2.1 & Gemini 3.6 Flash

Two videos watched end-to-end. What actually matters for our stack — and the one number that settles a question we were about to re-litigate.

Sources · Tech2WiLD — “Is This the BEST Local Model for One DGX Spark?” (12:11) · Fahd Mirza — “Gemini 3.6 Flash — Lost in the Australian Outback” (12:55)

★ The gold, in one line

An independent Tool-Eval Bench run put our exact quant — unsloth/Qwen3.6-35B-A3B-NVFP4-Fast — at 86.1, beating Poolside’s brand-new Laguna S 2.1 at 85.6. We don’t need to chase it. And the same video hands us the more valuable finding: thinking mode is a dial, not an upgrade — switching it on cost 12.3 points of agentic reliability and ran 30% slower.

That second one is directly actionable on Omnigent, VBL, and every tool-calling loop we run.

1 Laguna S 2.1 — Poolside’s 118B-A8B

Tech2WiLD · 12:11 · benchmarked on 4× RTX 3090 and DGX Spark

Poolside shipped the big sibling to their 33B-A3B “XS”: 118B total parameters, 8B active, MoE, built for long-horizon agentic and long-context coding work. The pitch is punching above its weight class — and the benchmark table backs it, against models 4× to 14× its size.

118B / 8B active
MoE — same active-param class as our 35B-A3B
250–256
tok/s on 4× RTX 3090, INT4 + DFlash, 200K ctx
30–40
tok/s on a single DGX Spark (peak decode 77)

Poolside’s own benchmarks — vs much larger models

BenchmarkLaguna S 2.1
118B-A8B
Tencent Hy3
256B-A20B
Nemotron 3 Ultra
500B-A50B
DeepSeek-V4-Pro-Max
1.6T-A64B
Inkling
Terminal-Bench 2.170.271.765.464.063.8
SWE-Bench Multilingual78.571.576.261.7

Beating a 1.6T model on SWE-Bench Multilingual with 8B active params is genuinely impressive. These are vendor numbers though — unverified on any hardware we own.

The head-to-head that matters to us we win

He ran Tool-Eval Bench — cross-trial summary comparison, agentic tool-calling workloads — between Laguna and the exact Qwen quant we already deploy:

Metricunsloth/Qwen3.6-35B-A3B-NVFP4-Fastpoolside/Laguna-S-2.1-NVFP4Δ
Mean score86.1 · winner85.6 · runner-up+0.5
Deployability (≥0.7)80 / 10071 / 100+9
Reliability (Pass²)68.7%63.7%+5.0
Safety warnings (max)12safer
Quality88 / 10090 / 100−2

Fewer points mean · lower variance · safer.” Laguna edges quality; we win everywhere that decides whether a model survives in an unattended agent loop.

“Thinking is a dial, not an upgrade” most transferable finding

Same model, same quant, thinking toggled OFF → ON:

AxisThinking OFFThinking ONΔ
Code7073+3
Instructions8692+6
Planning6570+5
Agentic99.392.6−6.7
Reliability on agents98.6%86.3%−12.3 pp
Latency (avg)5.9 s7.8 s30% slower
Overall (weighted)86.586.0−0.5
The rule it produces

Thinking ON to write code. Thinking OFF to run agents. Turning it on improves how a model reasons about a problem and degrades how reliably it executes a tool chain — and you pay 30% latency for the privilege.

DFlash is the real multiplier

Laguna ships with a DFlash draft model — install it plain, or install it with DFlash. On the same 4× 3090 box:

250 tok/s
With DFlash · 200K context · peak 256
80–90 tok/s
Without DFlash · but ~400K KV pool

That’s roughly 3× throughput for half the KV pool — the draft model eats the context budget. He published the config as a repo: “Laguna S 2.1 INT4 + DFlash at 200K on 4× RTX 3090.”

Single-Spark reality

Concurrency sweep on one DGX Spark, 1000-token generations:

StreamsTTFTServer tok/sDecode tok/s
×1368 ms35.0 (peak 117)41.0
×2526 ms58.4 (peak 175)45.7
×4578 ms56.2 (peak 198)27.5
×6742 ms111.6 (peak 333)29.2
×8736 ms78.0 (peak 408)26.2
His verdict

This is not the model for me — it’s not replacing my Qwen 35B-A3B and 27B.” It excels at coding and long-context building; for agentic work he stays on 35B-A3B. Running it costs him all four GPUs — losing concurrency, KV pool, and multimodality — to gain a smarter model he doesn’t need in the middle of a stack that already has GLM.

His hybrid stack, for reference: Fable/Opus on top builds → GLM 5.2 in the middle executes and supervises → 27B + 35B-A3B as workers. Notably close to ours.

2 Gemini 3.6 Flash — the agentic-completeness gap

Fahd Mirza · 12:55 · live agent test + multimodal sweep

Google dropped three Flash models at once. Only one is the story:

3.5 Flash-Lite3.6 Flash tested3.5 Flash Cyber
PositioningSpeed at scaleWorkhorse — better quality, lower costVuln detection in CodeMender
Input / Output$0.50 / $2.50 per M$1.50 / $7.50 per Mnot public
Speed350 tok/s17% fewer tokens than 3.5 Flash
Context1M1M
AccessPublicPublicTrusted gov partners only
BenchmarksTerminal-Bench 2.1 54%
OSWorld-Verified 74%
SWE-Bench Pro 54.2%
DeepSWE 49%
OSWorld-Verified 83%
MLE-Bench 63.9%
CyberGym frontier-competitive

His read: Google is upgrading against its own prior models, not against the field. 3.6 Flash beats 3.5 Flash on coding, computer use, and MLE-Bench — while using fewer tokens per task. It explicitly does not claim to beat 3.1 Pro on the hardest long-context reasoning; that gap stands. Flash-Lite is the sleeper — it beats the older 3 Flash on SWE-Bench Pro and OSWorld while being cheaper and faster.

The live test — and the exact failure

He built a FastAPI seaport freight tracker with a real bug planted in it: a vessel with a 9.8 m draft gets silently approved into a 9.5 m berth instead of rejected — in the real world, the ship runs aground. He handed the whole repo to 3.6 Flash through a Hermes agent with one goal: find the bug, fix it, start the server, and verify the fix yourself.

✓ Found & fixed
No hints, no pointer to the file. Came back “fairly quickly.” The fix was correct — the berth assignment now rejects.
✗ Left it dirty
Did not restart the backend, did not clean the stale DB file. He did both by hand before the fix was visible.

His pointed comparison: Gemini K3 and Fable restarted those backends themselves. 3.6 Flash didn’t. The intelligence was there; the follow-through wasn’t.

Multimodal sweep

His verdict

“Quite good from Google, but I don’t think in terms of very high quality coding they match Fable 5 or Opus — still not a bad result at all.”

3 What we do with it

① Don’t swap the 35B-A3B. Log it as a standing verdict.

An independent tool-calling bench has unsloth/Qwen3.6-35B-A3B-NVFP4 beating Laguna 86.1 to 85.6, with better deployability (+9), better reliability (+5 pp), and fewer safety warnings. Laguna stays ⚠ claimed in the Model Diary until measured on our own iron — and this is exactly the “should X replace our core local?” question the Standing Verdicts table exists to stop us re-arguing.

② Turn thinking OFF on agentic paths. highest value

−12.3 pp reliability on agents, −6.7 agentic score, +30% latency — for +3 on code. Every place we run a tool-calling loop (Omnigent dispatcher, VBL voice agent, fleet-rpc job flows, the vb-broker fan-out) should default thinking off and switch it on only for code generation. This costs nothing to test and is the single most portable finding across both videos.

③ DFlash is the multiplier worth measuring — on our iron.

80–90 → 250 tok/s from the draft model alone, same GPUs. We already run Buun-DFlash on our orchestrator. Worth measuring that ~3× ratio on our 96 GB box’s 96 GB and on our 24 GB laptop-GPU nodes, and deciding deliberately whether we’d trade half the KV pool for it per workload. Speculative decode may be a bigger lever for us than any model swap this month.

④ Laguna on a Spark isn’t compelling for us.

30–40 tok/s on a single Spark, and it wants the whole box. We already have DeepSeek-V4-Flash measured at 59.5 tok/s on one of ours. The 4×3090 path is where Laguna shines, and that isn’t a shape we have.

⑤ Treat model agentic-completeness as a harness problem, not a model problem.

3.6 Flash fixed the bug and left the environment dirty. Any model we wire into an unattended loop will do some version of this. The restart / re-verify / clean-state step belongs to our harness, asserted independently — never trusted to the model’s own “I’m done.”

⑥ Flash-Lite is the only cloud line item here worth a second look.

$0.50/M in and 350 tok/s at 1M context is genuinely cheap for high-volume classification or extraction. But our own hardware is free — this only earns a slot where scale or latency beats what we can serve ourselves. Not a default.

Provenance: every number on this page is from the two videos — vendor benchmark tables, or measured by the presenters on their hardware (4× RTX 3090, DGX Spark). None of it has been measured on fleet iron. Per Model Diary law, treat all of it as ⚠ claimed until we run the same-prompt harness ourselves.