Gemini 3.8 Flash

“I'd say it's about equal to luna, but FAST. And terra is definitely more intelligent. Gemini 3.8 flash still has that old gemini problem of often just going for the quick surface solution.”

“It can disentangle and understand a 100,000-word mystery novel with a chaotic timeline and narrative tricks, something none of the other models I tried could do.”

“Gemini 3.8 flash is actually pretty good now.”

“I did gem 3.8 flash on medium I feel its the perfect one for that because sometimes it can over think”

“It still drools or slacks off when the process gets long, so I gave up on it early and use it for coding instead.”

“I think 3.7 and 3.8 are a significant improvement over 3.1.”

“This one doesn't perform as well as GPT.”

“Same model, one version of flash, but the pro suddenly became very cold and formal. Also doesn't like speaking in a 1, 2, 3 style anymore. Analogies have all disappeared.”

“Gemini confidently ignores instructions. I thought things would get better with 3.8, but somehow it feels even worse. Even when I explicitly tell it something like "just tell me", it goes straight into editing files.”

“DeepSWE v1.1, 3.8 Flash scored 73.7%, an improvement of over 8 percentage points from 3.7 Flash's 65.3%, with Ars Technica reporting it has topped the benchmark. On Terminal-bench 2.1, which focuses on agent-side coding, it achieved 89.4%, surpassing a range of flagship models.”

“Luna sucks. Use Gemini 3.8 flash ffs”

“You say it's not good? It has a few yuan 18-month subscription on the second-hand market, I've been using it for at least 3 weeks and already gotten my money's worth, its capability is definitely above the value of 10 RMB. You say it's good? 30% of the time it needs another model to wipe its butt.”

“3.8 flash is definitely good for coding”

“gemini 3.8 flash is genuinely excellent at coding. It's the instruction following and making wrong assumptions that suck and the annoying thing where you give it a plan and it comes back with an inferior plan. But one thing i've noticed if”

“Gemini 3.8 Flash (which is an _excellent_ model especially for its price and TPS!!”

“Not as good as Doubao”

“Gemini 3.8 flash is amazing for a small model”

“Son raisonnement est merdique comparé à gemma 4 e2b en qat”

“3.8 flash is benchmaxxed to hell, garbage model”

“Still top-tier at benchmark grinding, top-tier at scores, actual performance is garbage”

“gemini 3.8 flash spent 2 minutes writing more than two screens worth, I couldn't be bothered to read what its main points were”

“Besides being fast, it doesn't seem to have other advantages”

“Ignoring prompts, agents.md and other file requirements, being smartass, resulting in output that looks pretty but is actually a pile of dog shit”

“Flash's peers are Sonnet/Terra; comparing it with Opus/Sol is really giving Google too much credit”

“Based on my experience, its improvement in coding is quite remarkable. Comparing 3.6 -> 3.7 from unusable to usable, I personally think 3.7 -> 3.8 can be considered from usable to good”

“Requires you to put effort into task decomposition and prompts, but compared to 3.7 which couldn't even execute well-made plans, 3.8 is much stronger”

“Don't know about the rest, but RP enthusiasts will love it, feels very alive”

“Deep SWE is single-task short programming, while Terminal Bench 4 is long-task programming averaging five to six hours. Flash model, Terminal Bench 4 basically won't be very high”

“gemini's output speed is 300t/s”