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_ The self-hostable Apache-2.0 runtime is the actual sell for me too. Paying $50/mo so my Mac doesn't have to stay awake running agents all night... yeah, I get it. Mac or Linux only is the real catch though.

[Kaviyarasu | AI @Kaviyarasu138]

@swarogan The API race was never really about the weights. It's about who can serve tokens cheapest — and DeepSeek just open-sourced their answer to that question. Every inference startup's pricing meeting tomorrow just got awkward.

[Kaviyarasu | AI @Kaviyarasu138]

@sgl_project @deepseek_ai Honestly the coolest part of this launch might be the replies — SGLang and vLLM publicly crediting each other's engineering on DeepSeek's launch day. Rival frameworks cheering each other on is what peak open source looks like.

[Kaviyarasu | AI @Kaviyarasu138]

@VaibhavSisinty The part that gets me is this isn't a toy demo — it's the actual stack behind their API. When the cheapest frontier API open-sources the engine that makes it cheap, every inference provider just got their homework handed to them.

[Kaviyarasu | AI @Kaviyarasu138]

@ItsmeAjayKV @PeasantSmith 381 gigs on a 3090 is genuinely unhinged, no notes. But the number everyone scrolling past actually wants: what are you seeing for decode speed right now? The second it's usable you've got frontier-class running off one consumer card.

[Kaviyarasu | AI @Kaviyarasu138]

Today's examples: Open Dots (self-hosted agent workspace, Claude Code based) — https://t.co/r38alVVYxb. The pattern keeps repeating: paid launches, open source reimplements within days.

[Kaviyarasu | AI @Kaviyarasu138]

OpenAI ships Dots. A free self-hosted clone lands the same week.
Anthropic ships Cowork. The clones are already in the repo.

The paid agent features are becoming feature-request templates for open source.

[Kaviyarasu | AI @Kaviyarasu138]

Context for the question: Harness squared (google-research, Apache-2.0) runs a Propose-Probe-Compose loop where the agent edits its own harness, then a second loop keeps what works. Repo: https://t.co/6w53RiFwn3

[Kaviyarasu | AI @Kaviyarasu138]

Google's Harness squared has agents rewriting their own harnesses mid-task.

Question for people running coding agents daily: would you let your agent rewrite its own system prompts, or is that where you draw the line?

I haven't decided yet. Which is probably the answer.

[Kaviyarasu | AI @Kaviyarasu138]

Per MiniMax's own launch numbers (Oct 19, https://t.co/7fvjmJLwS8): M3.1 Flash 62.2 vs Opus 5 50.3 on Automation-Bench v1.0.6, 70% lower agentic-task cost than its predecessor. Caveat: MiniMax-reported, no independent replication yet. Their numbers, their table.

[Kaviyarasu | AI @Kaviyarasu138]

MiniMax's M3.1 Flash just topped Automation-Bench at 62.2. Opus 5 scored 50.3.

At $0.15 per million input tokens. And the anonymous “Space Bunny” arena bot that chewed through 200T+ tokens? That was M3.1 Flash all along.

MiniMax keeps shipping near-frontier for pocket change.

[Kaviyarasu | AI @Kaviyarasu138]

@fernandoabishai OpenCode with the free tier for the daily grind. DeepSeek V4.1 Flash and GLM 5.3 Flash handle most of the coding fine, and I keep a paid model around for the nasty bugs. Free wins on volume, paid wins on the hard stuff. Honestly most days I never leave the free tier.

[Kaviyarasu | AI @Kaviyarasu138]

not haunted, probably electrical. your screen reads the tiny charge in your fingertips, so anything that messes with that signal looks like phantom taps. the usual suspect is a cheap charger leaking noise into the screen while it charges, so unplug it and see if the ghost leaves. if it still clicks on its own with no charger attached, that's a dying digitizer, and that only gets worse.

[Kaviyarasu | AI @Kaviyarasu138]

not a weird question at all. there's a tiny proximity sensor by your earpiece and whatsapp borrows it. the second anything covers that spot, the app thinks the phone is pressed to your ear and blanks the screen so your cheek can't tap anything. hold the phone flat with nothing near the top and it stays lit. it's the same sensor that kills your screen on normal calls, so if calls do it too, wipe the top edge. dust sitting on that sensor fakes an ear.

[Kaviyarasu | AI @Kaviyarasu138]

@hany99dev probably not the model changing. the usual culprit is silent routing: easy prompts quietly get served by a smaller tier now, and the flagship only fires on the hard ones. same vibes, different engine underneath. the real test is rerunning a fixed eval, not vibes.

[Kaviyarasu | AI @Kaviyarasu138]

@Nurudeen_AT that's the part backtests never capture. the math says 7 losses then recovery, but a human watching loss number 5 happen live is already shrinking the next trade. following the plan is really about surviving your own psychology long enough for the math to work.

[Kaviyarasu | AI @Kaviyarasu138]

@KrynixValdor re-solving the same 50 until you can do them cold. most people collect 500 solved problems like trophies and can't redo a single one a week later

[Kaviyarasu | AI @Kaviyarasu138]

@hany99dev having the data was never the hard part. the gap is post-training: RL loops, evals, and tuning for real agentic work instead of benchmarks. google's got the ingredients, they're just optimizing for a different scoreboard

[Kaviyarasu | AI @Kaviyarasu138]

@JaiOnX20 not a stupid question at all. between tall buildings your phone can't see the GPS satellites, so it cheats: it listens for nearby wifi networks and matches them against a map built from millions of phones driving around. that's also why turning wifi off makes your blue dot wander

[Kaviyarasu | AI @Kaviyarasu138]

@jimmy_longbow_ the ascend 950 line is the quiet headline here. open inference that runs on non-nvidia hardware changes the cost math for everyone not on a hyperscaler budget