Mauricio Gallmur

High-ticket closer → AI Builder.
I build, ship, break, measure & document AI systems & products.
One brand. Real work. No hype.

[Mauricio Gallmur @MGallmur]

DeepSeek V4 Pro GA estimates:

Terminal Bench 2.1 → 88-92
DeepSWE → 55-65
Cybergym → 78-85

Based on the exact jump Flash just made with post-training. Pro has 3.8× more active parameters.

This might actually challenge GPT-5.6 Sol. The open-weight king is loading…

[Mauricio Gallmur @MGallmur]

@saddab_0 Let’s do it

[Mauricio Gallmur @MGallmur]

I caught X search inventing three plausible Hermes features this week. Names, architecture, even install instructions. None existed in the official release notes.

Broad search is useful for discovery, but I’d never let agent-generated research reach production without a primary-source check.

What verification gate do you use before your agents act on research?

[Mauricio Gallmur @MGallmur]

@nikssocial Public affirmations turn your goals into a real commitment to yourself

[Mauricio Gallmur @MGallmur]

@Pokee_AI 10M real context on a single GPU is the only part that matters.
Going to test actual multi-step agent performance, not the bench numbers.

[Mauricio Gallmur @MGallmur]

10M context on a single RTX 4090 + $0.15/M input?

If the RULER numbers and agentic claims hold in real workflows (not just the bench), this becomes interesting for long-horizon agents.

Still closed source for now. Going to test the actual task cost, not the pricing page.

Anyone already running it?

[Pokee AI @Pokee_AI]

Releasing Pokee-Isaac 28B — the world’s first real 10M-token context frontier-class agentic model, deployable on a single GPU (starting from RTX 4090 or equivalent).

New proprietary non-decoder-only architecture:
• 93.3% RULER at 10M tokens
• Up to 137K tokens/s prefill on one B200 with 10M-token context
• Leads BFCL v4 and τ³-bench in our evaluation
• Lowest combined attack success rate among evaluated models on DTAP security red-teaming benchmark

Pricing and deployment:
💰 $0.15/M input · $1/M output
🔒 Deploy in your VPC, on-premises, or on-device, with Day-0 support for @vllm_project and @sgl_project

Technical blog: https://t.co/nFqaYBlcQP
Technical report: https://t.co/XDOoZxpgJx
API: https://t.co/KYj8fOOjNS

[Mauricio Gallmur @MGallmur]

@diegohaz You’re right that’s why gpt luna is a trap

https://t.co/HGq5s0WTke

[Mauricio Gallmur @MGallmur]

GPT 5.6 Luna could be a trap!

At $0.20/M input, Luna is cheap enough to enter the default routing conversation.

But token price is only the admission fee. If it needs more retries, verification, or human cleanup, the cheap lane gets expensive fast.

Route on total task cost, not the pricing page.

[Mauricio Gallmur @MGallmur]

@itzs_julien When it didn’t match my expectations

[Mauricio Gallmur @MGallmur]

@i_mika_el The favor of consistency and results is the greatest

[Mauricio Gallmur @MGallmur]

@quxiaoyin Who else here is a Hermes agent heavy user with DSv4 flash?

[Mauricio Gallmur @MGallmur]

@sri9s Time has come!

[Mauricio Gallmur @MGallmur]

@Yuhao_Guo1018 Darío

[Mauricio Gallmur @MGallmur]

@mark_k we're just warming up for what's to come

[Mauricio Gallmur @MGallmur]

@SherryYanJiang Some rumors suggest that Fable 5.1 is imminent.

[Mauricio Gallmur @MGallmur]

Update https://t.co/gGRI4l8obv

[Watcher.Guru @WatcherGuru]

UPDATE: Telegram has been added back to the Apple App Store.

[Mauricio Gallmur @MGallmur]

Telegram just got pulled from the Apple App Store worldwide 🚨

No official reason yet. Existing installs still work. New iOS downloads are blocked.

For anyone running agents through Telegram (Hermes and similar harnesses):

The agent itself is fine.
Telegram is just the delivery channel.

If you already have the app → keep using it.
If you don’t → web version, desktop, or switch the gateway.

This is a good reminder: never make a single messaging app a single point of failure for production agents.

Build the system so the harness survives the client.

[Watcher.Guru @WatcherGuru]

JUST IN: Telegram removed from the Apple App Store worldwide.

[Mauricio Gallmur @MGallmur]

@emeeliojohann it's back https://t.co/gGRI4l8obv

[Watcher.Guru @WatcherGuru]

UPDATE: Telegram has been added back to the Apple App Store.

[Mauricio Gallmur @MGallmur]

@emeeliojohann Yes, it is! It'll probably be back soon for new users.

[Mauricio Gallmur @MGallmur]

hey @ArtificialAnlys what happened with your benchmarks

[Mauricio Gallmur @MGallmur]

Qwen3.8 Max improved its agentic scores, but the operational picture is mixed.

Artificial Analysis measured it at $1.76 per task; roughly 2× Kimi K3 and 3× GLM-5.2. It also used 50% more output tokens than its predecessor, while hallucinations increased on one evaluation.

I haven’t tested it, so I don’t have a verdict.

Benchmarks earn a model a place in the evaluation queue, not in production.

[Mauricio Gallmur @MGallmur]

@humfhuang The algo is popping your account, I guess. Well done

[Mauricio Gallmur @MGallmur]

@araseb_ 100% agree

[Mauricio Gallmur @MGallmur]

@krishnapro_ I don't like linkedin, so X is my go to