Run GLM-5.2 Locally

In today’s newsletter:

• Apodex: The World's First Self-Evolving Heavy-Duty Solver

• Firecrawl Prometheus: Prompt an Agent to Collect Web Data

• Unsloth: GLM-5.2 Can Now Run Locally

Reading time: 5 minutes.

The World’s First Self-Evolving Heavy-Duty Solver

Getting an answer from a research agent is easy. Knowing whether to trust that answer is not.

Most agents retrieve, synthesize, and deliver a conclusion. They don't trace claims back to sources. They don't flag when a premise is wrong. You get an answer with no way to verify it.

Apodex is built around a different idea: finding the answer is only half the job. The other half is proving it's right.

Here's how it works:

When you submit a query, a main agent breaks it down and dispatches specialized sub-agents to research in parallel, each running with its own context and tool set.

As reports come in, a dedicated verification team reviews them: a conflict reviewer checks where findings disagree, a fact checker grounds individual claims, and a draft-report reviewer audits the full answer before anything is committed.

The solver and verifier never share context. They use different tools. A verifier that has seen the solver's reasoning will rationalize. One that hasn't, won't.

If a claim doesn't hold up, it gets cut before delivery.

What this looks like in practice

Take this question: "Across the past eight quarters, has Adidas management systematically underestimated FX impact relative to its own guidance?"

The question has a hidden flaw. Adidas doesn't publish quarterly FX guidance. There's no baseline to compare against. A research agent that skips verification will invent one.

Apodex caught it before answering. It flagged that the premise was incorrect, explained that the comparison the question assumes doesn't exist in any public filing.

Key features:

Solver-verifier separation: The verifier runs with its own context and tools. It hasn't seen the solver's reasoning. It can disagree.

Auditable trail: Every finding and verdict is recorded. Any conclusion can be traced back, challenged, or revised.

AgentOS runtime: The layer that runs the agent team. Handles scheduling, model and tool routing, checkpoints, and cost accounting.

Open-weight models: Apodex-1.0-mini (35B-A3B) and the 0.8B, 2B, and 4B variants. The 4B outperforms every open-source 30B-class model on both BrowseComp and BrowseComp-ZH.

Run the benchmarks yourself

The evaluation harness used to produce these numbers is open source. AgentHarness supports public deep-research benchmarks, including BrowseComp, DeepSearchQA, HLE-Text, and FrontierScience.

If you're building a research agent and want to know where it actually stands, this is the harness to run it through.

Check out AgentHarness on GitHub →

Firecrawl Prometheus: Prompt an Agent to Collect Web Data

Collecting web data usually means writing scrapers, handling page changes, and maintaining brittle extraction code.

Even with tools like Firecrawl, someone still has to figure out how to collect the data and keep the pipeline working as websites evolve.

Prometheus solves that.

How Prometheus Solves It

Prometheus is a web data agent from Firecrawl. Describe the data you need in plain English, and it generates a Firecrawl collector to gather it.

You tell Prometheus what data you want. It generates a Firecrawl SDK collector and returns both the code and sample data. The collector is yours to run, modify, and version however you like.

Trying It Out

To see how Prometheus works, we asked it to monitor the pricing pages of OpenAI, Anthropic, Google Gemini, and Cohere.

Prompt:

Prometheus generated a Firecrawl collector to gather the data and returned the corresponding code and output. Instead of manually building and maintaining scrapers for each provider, the agent handled the implementation from a single prompt.

What This Enables

• Competitor monitoring

• Product catalog tracking

• Lead generation workflows

• Market research pipelines

• Automated web data collection without maintaining scrapers

We're starting to see a shift from AI helping developers write scrapers to AI owning the entire data collection workflow.

Instead of building extraction pipelines by hand, developers describe the outcome they want, and an agent generates, tests, deploys, and maintains the implementation.

Try out Prometheus →

GLM-5.2 Can Now Run Locally

GLM-5.2 is Z.ai's latest open model: 744B parameters, 40B active via MoE routing, and a 1M token context window. On benchmarks, it sits on par with Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro across reasoning, coding, and agentic tasks.

The problem: the full model requires 1.65TB of storage. That ruled out local inference for almost everyone.

Unsloth's Dynamic 2-bit GGUF changes that. It shrinks GLM-5.2 to 238GB while retaining roughly 82% accuracy. That's an 86% reduction in size, enough to fit the entire model on a 256GB unified memory Mac or a combined RAM and VRAM setup.

Memory Requirements by Quantization

The 2-bit quant works well on a 256GB Mac or a setup combining a single 24GB GPU with 256GB of system RAM using MoE offloading.

Three Thinking Modes

GLM-5.2 has three inference modes: non-thinking, High thinking, and Max thinking. Max thinking is designed for complex reasoning and agentic tasks.

In Unsloth Studio, you can toggle between modes from the UI.

How to Run It

Unsloth Studio (easiest)

Open http://127.0.0.1:8888, search for GLM-5.2, and download. Unsloth Studio automatically handles RAM offloading and multi-GPU detection.

llama.cpp

Or download manually first:

Benchmark Highlights

On the benchmarks where it competes most directly with frontier closed models:

• AIME 2026: 99.2 (vs Claude Opus 4.8's 95.7)

• IMOAnswerBench: 91.0 (vs Claude Opus 4.8's 83.5)

• SWE-bench Pro: 62.1 (vs Claude Opus 4.8's 69.2)

• MCP-Atlas: 76.8 (vs Claude Opus 4.8's 77.8)

GLM-5.2 leads on math reasoning. Claude leads on coding and agentic benchmarks. Across the full benchmark suite, they trade wins, which for an open model is a significant result.

Full setup guide → 

That’s all for today. Thank you for reading today’s edition. See you in the next issue with more AI Engineering insights.

PS: We curate this AI Engineering content for free, and your support means everything. If you find value in what you read, consider sharing it with a friend or two.

Your feedback is valuable: If there’s a topic you’re stuck on or curious about, reply to this email. We’re building this for you, and your feedback helps shape what we send.

WORK WITH US

Looking to promote your company, product, or service to 200K+ AI developers? Get in touch today by replying to this email.