Kaviyarasu | AI

Breaking down AI so you don't have to. Free models, tools and tricks daily.

[Kaviyarasu | AI @Kaviyarasu138]

Tokens processed in the last 30 days on OpenRouter:

DeepSeek V4.1 Flash → 32.8 trillion
GLM 5.3 Flash → 12.2 trillion
Step 5 Preview → 11.5 trillion
MiMo V2.6 Flash → 9.94 trillion
Hy4 Preview → 7.55 trillion

The most-used AI models on Earth are cheap, open, and mostly Chinese. The leaderboard nobody posts about.

[Kaviyarasu | AI @Kaviyarasu138]

@jimmy_longbow_ Honestly? Long runs are exactly where I'd still rent. Everything else lives on free models through OpenCode on my machine and they hold up fine for normal work. But a three-hour autonomous run is where the reliability gap shows. Rent the hard 20%, run the rest local.

[Kaviyarasu | AI @Kaviyarasu138]

@olive_jy_song The stealth launch was the real proof of the RL work honestly. 12 days as the #1 model on OpenCode before anyone knew what it was, 106T tokens of people just choosing it in real sessions. Benchmarks are the announcement, usage was the evaluation.

[Kaviyarasu | AI @Kaviyarasu138]

@jenzhuscott Getting dethroned in 2 days stings a lot less when your model is free. DeepSeek's real lead was never the benchmark crown, it was making everyone else race to the bottom on price. MiniMax won a battle, DeepSeek already won the war.

[Kaviyarasu | AI @Kaviyarasu138]

@mehulmpt The funniest part is they've shipped flash models so fast the platforms can't keep up. OpenCode still serves the old version while 1017 is sitting there free. The jump from the old build to this one is genuinely huge.

[Kaviyarasu | AI @Kaviyarasu138]

How to get it: https://t.co/0rTahmYeFz, grab an API key, model id is deepseek-v4.1-flash-uncensored, OpenAI-compatible endpoint at https://t.co/rLu3o1hEt6. One honest note: their catalog lists the standard rate for this model at $0.30 in / $1.20 out per million tokens, while the model page currently shows it free for a limited time, so treat the free window as a launch offer and verify at the console before you build on it. Their gateway itself is open source with 0% markup on routed tokens

[Kaviyarasu | AI @Kaviyarasu138]

DeepSeek V4.1 Flash with the refusals removed. Free. Right now

Experiential Labs took DeepSeek's V4.1 Flash, stripped the refusals out of the weights themselves, and put it on their own GPUs. The model page says free for a limited time

1M token context. Thinking on by default. Tools and reasoning fully supported. And the weights are public too, dealignai on Hugging Face, so this is not a demo you have to trust, you can run the same weights yourself

Free AND uncensored AND a million tokens of context. That combination basically did not exist last month

[Kaviyarasu | AI @Kaviyarasu138]

@notjazii The overthinking is the real tax though. 200+ tokens per second means nothing if it spends 400 tokens thinking before every answer. DeepSeek flash gets away with it because it's free, M3.1 at $0.60 out has to justify every token.

[Kaviyarasu | AI @Kaviyarasu138]

@dummerspast39 It was the #1 model on OpenCode for 12 days as a stealth launch though. 106T tokens of people actually choosing it in real sessions. The 42 might be debatable but the usage isn't.

[Kaviyarasu | AI @Kaviyarasu138]

@ChrisGPT Benchmaxing or not, this is the first chart that measures what agent work actually costs. Score per dollar per finished task is the only leaderboard that matters and now everyone else has to beat it. The frontier just became a price war.

[Kaviyarasu | AI @Kaviyarasu138]

@arena @MiniMax_AI WebDev arena is one shot landing pages though. Nobody scores whether the model survives a three hour refactor session without breaking the repo. Flash models keep winning the wrong game because the right one isn't measured yet.

[Kaviyarasu | AI @Kaviyarasu138]

@DeItaone The actually interesting part isn't reading intentions better, it's asking first. Every agent framework is racing toward more autonomy and Anthropic's move is a model with better restraint. The models that ask before nuking your codebase win.

[Kaviyarasu | AI @Kaviyarasu138]

@LuminaBench It doesn't need to beat Opus, it needs to survive a real agent session. The frontier race keeps optimizing benchmark numbers while the models still fall apart at step 40 of actual tool loops. Ship reliability, not a higher score.

[Kaviyarasu | AI @Kaviyarasu138]

@opencode The wildest part is it was #1 for 12 days while nobody knew what it was. 106T tokens of free usage and the market just quietly voted. Stealth launch plus free is honestly the best model evaluation we've got.

[Kaviyarasu | AI @Kaviyarasu138]

@Mr_nexpo 2.5 trillion parameters is the least interesting part of this. The models winning my agent sessions right now are flash sized and free to run. If Grok needs 2.5T to compete, that's not a flex, that's overhead.

[Kaviyarasu | AI @Kaviyarasu138]

@yuhasbeentaken Same experience here. The benchmark numbers are fine but they never test what actually breaks in a real agent session. M3.1 flash stalls on tool calls in OpenCode while DeepSeek flash at $0 just keeps running. Cheap per token means nothing if the session dies at step 40.

[Kaviyarasu | AI @Kaviyarasu138]

@bindureddy Half of Luna still isn't free though. My OpenCode runs are already on DeepSeek flash at $0, and the thing that would actually win me back isn't cheaper tokens, it's a small model whose tool calls don't fall apart at step 40. Cheap and dumb loses to free and reliable every time.

[Kaviyarasu | AI @Kaviyarasu138]

@itsPolloAI Funny thing is this isn't even a fair fight. Opus 5.5 costs real money per run and DeepSeek flash is basically free. Flash doesn't need to be prettier, it just needs to be close enough at 1% of the cost. And it's getting there.

[Kaviyarasu | AI @Kaviyarasu138]

@StocksToTrade Classic Sunday night short cover into the weekly open. Question is whether Monday's killzones push this into the FVG sitting above or if it's just a relief pop before the real sellers step back in.

[Kaviyarasu | AI @Kaviyarasu138]

@StatsWire Everyone's bracing for the 3.8 promo to expire, but the bar moved while they weren't looking. DeepSeek V4.1 Flash runs my agent sessions for free and M3.1 Flash is $0.15 in. Google can't just lift flash prices back up anymore, 4 Flash has to earn that premium.

[Kaviyarasu | AI @Kaviyarasu138]

@LuminaBench Benchmarks never measure what actually matters for agent work though. A 42 that survives a three hour agent session beats a 45 that falls apart at step 40. Reliability under tool loops is the real benchmark and nobody's charting it.