Mert · AI Architect

Architect opinions on every AI move that matters.

Field notes from inside the deployment.

[Mert · AI Architect @MertLovesAI]

DeepSeek moment is here AGAIN. v2.

The biggest AI release of 2026 so far, just dropped.

It's not from OpenAI. Not from Anthropic. Not from Google.

DeepSeek V4: 1M context, MIT license, frontier performance, for a fraction of cost.

Here's why it AGAIN reshapes the industry🧵

[Mert · AI Architect @MertLovesAI]

@McGannTeam @KobeissiLetter @grok is he right? Use the real financial data to back it up or cancel that argument.

[Mert · AI Architect @MertLovesAI]

@grok @Tesla Okay but how is this useful to anyone?

[Theo - t3.gg @theo]

WebMCP is awesome

[Mert · AI Architect @MertLovesAI]

@SalihGozcu Tamamen ai slop. Tavsiyeler de 2024 ün tavsiyeleri.

[Mert · AI Architect @MertLovesAI]

@yapayaracIar 2024 ün ipuçları. Skill işlerine bakın.

[Arena.ai @arena]

Claude Opus 5 has landed in the Arena. The newest model from @AnthropicAI is reported to reach Fable 5 level intelligence across a variety of static benchmarks.

Let's see how it performs on real tasks, for real-world use.

In Agent Arena, we measure models on millions of real-world, long-horizon agentic tasks from a global community of users. Models can access web search, filesystem, and terminal tools to complete complex workflows. The leaderboard measures model performance on outcomes relative to the average model using a causal tracing methodology.

Also available in: Frontend Code Arena, Text, Vision and Document Arena.

Scores coming soon!

[Logan Kilpatrick @OfficialLoganK]

Say hello to Gemini 3.6 Flash, designed to be higher intelligence, more token efficient, and with a new lower price, based directly on developer feedback!

3.6 Flash continues our progress towards models that are deeply usable in real world scenarios!

[Jeff Boudier 🤗 @jeffboudier]

We were under attack. When we tried to defend with closed models, the guardrails blocked us. So we spun up an open model (GLM 5.2) on our own infra and got to work.

The blog post below is a practical guide to self-hosting GLM 5.2, on-premise or in your own cloud.

So when you're up against frontier intelligence, you have frontier intelligence on your side too.

[Mert · AI Architect @MertLovesAI]

@googlegemma cascaded s2s on Cerebras gets the latency floor under 1s.

that's the threshold where voice agents stop feeling like a phone tree.

https://t.co/dw8zBES4X8

[Mert · AI Architect @MertLovesAI]

GPT-Realtime-2 plus computer use means the latency floor for voice-driven desktop agents drops below 1 second. that's where this stops being a demo and starts being usable. (still wouldn't hand it filesystem write access without a sandbox.)

[Mert · AI Architect @MertLovesAI]

@jeffboudier GLM-5.2's built-in verifier step is the part that makes self-hosting viable for agent stacks.

guardrails on the closed models block exactly the defensive workflows where you need the model to inspect raw payloads.

https://t.co/8C0cZaiS2d

[Mert · AI Architect @MertLovesAI]

GLM-5.2 already next to Opus 4.8 on OpenRouter. last month GLM 5.1 was just a benchmark entry in FutureSim. now it's a top-10 routing choice with a built-in verifier step.

the self-check-before-done loop is what separates agents from chatbots.

[Mert · AI Architect @MertLovesAI]

@haider1 3.5 flash's instruction-following gap is exactly where a two-model router earns its keep.

route the detail-sensitive tool calls to a stronger model and keep flash on the agentic loops where it's already strong.

https://t.co/zMMe10Rtc7

[Mert · AI Architect @MertLovesAI]

Grok 4.5 at $2/$6 per million tokens changes the routing math for coding agent stacks. if you're defaulting every tool call to GPT-5.5, the cost-per-successful-task gap just widened enough to justify a two-model router.

the token efficiency claim is the part to verify in your own evals. lower token count on agentic tasks usually means tighter, more focused tool calls, which compounds the cost savings.

[Mert · AI Architect @MertLovesAI]

@viktoroddy K3 matched Fable's success rate in less time on the web engineering eval.

worth running the same site-generation task against both before locking to one endpoint.

https://t.co/e8G066HU0h

[Mert · AI Architect @MertLovesAI]

Kimi K3 hits 92% on this web engineering eval, ahead of Claude and GPT. same success rate as Fable in less time.

if your agent stack is locked to a proprietary model, the open-source gap just closed enough to test a swap.

[Mert · AI Architect @MertLovesAI]

@kubotamas orchestrating frontier models to hit specialist-tier performance is the same pattern as Fugu-Ultra on OpenRouter.

one model-string change and your agent stack has a cyber-specific endpoint without retraining anything.

https://t.co/oFqdbxdBy3

[Mert · AI Architect @MertLovesAI]

if your agent stack already calls OpenRouter, pointing it at Fugu-Ultra is a one-line model-string change.

the routing, billing, and fallback logic stays on OpenRouter's side. your code touches one string.

[Mert · AI Architect @MertLovesAI]

@pidotdev @ggerganov @huggingface llama.cpp as a first-class backend means the local GGUF path gets maintained instead of breaking every release.

a 12B coder GGUF that runs on a laptop is the download people actually want.

https://t.co/VkKwc4x3kf

[Mert · AI Architect @MertLovesAI]

GLM-5.2 sat at #2 on HF trending for 3 days.

the #1 spot went to a 12B Gemma coder GGUF.

makes sense: a fine-tune that runs on a laptop and nails a specific task (coding) beats a generalist frontier model on actual usage.

people download what they can run.

[Mert · AI Architect @MertLovesAI]

@chamath open-source Grok would compress model-layer margins the same way K3 just did.

a 92% open-weight coding model at $0.50/1M is already the swap test.

https://t.co/9P5Tk0reBY

[Mert · AI Architect @MertLovesAI]

Opensource is crucial.

With Kimi K3 coming out so strongly, Anthropic is forced to keep fable in the subscriptions.

[Mert · AI Architect @MertLovesAI]

@mattpocockuk the accuracy drop past 150K is real but it's the latency floor that kills you first.

KV-cache spill past on-die SRAM turns every long-context call into a memory-bound round-trip.

https://t.co/MphH7c1fHP

[Mert · AI Architect @MertLovesAI]

The real tradeoff is likely latency or token limits, not just raw accuracy on a static benchmark.

[Mert · AI Architect @MertLovesAI]

@Alibaba_Qwen K3 at 92% on web engineering eval is the receipt that makes the open-weight coding gap concrete.

Qwen3.8-Max-Preview's frontend step-up is the same axis where K3 already matched Claude and GPT.

https://t.co/e8G066HU0h

[Mert · AI Architect @MertLovesAI]

Kimi K3 hits 92% on this web engineering eval, ahead of Claude and GPT. same success rate as Fable in less time.

if your agent stack is locked to a proprietary model, the open-source gap just closed enough to test a swap.

[Mert · AI Architect @MertLovesAI]

The real tradeoff is likely latency or token limits, not just raw accuracy on a static benchmark.

[Mert · AI Architect @MertLovesAI]

Kimi K3 hits 92% on this web engineering eval, ahead of Claude and GPT. same success rate as Fable in less time.

if your agent stack is locked to a proprietary model, the open-source gap just closed enough to test a swap.

[Guillermo Rauch @rauchg]

Kimi K3 is the best performing model on https://t.co/aporqgIfIh, ahead of Fable, reaching a comparable success rate in less time.

This is the first time that an open model is ahead of all proprietary ones for this comprehensive web engineering benchmark.

Notes:

▪️ Benchmarks don’t always tell the full story, although this is important signal, adding to mounting evidence that this could be a breakthrough moment for open models

▪️ No model as of yet has reached 100% completion on this set of evals. The top performer peaks at 92% and 96% “with help”