📆 ThursdAI - Dec 4, 2025 - DeepSeek V3.2 Goes Gold Medal, Mistral Returns to Apache 2.0, OpenAI Hits Code Red, and US-Trained MOEs Are Back!

Hey yall, Alex here 🫡

Welcome to the first ThursdAI of December! Snow is falling in Colorado, and AI releases are falling even harder. This week was genuinely one of those “drink from the firehose” weeks where every time I refreshed my timeline, another massive release had dropped.

We kicked off the show asking our co-hosts for their top AI pick of the week, and the answers were all over the map: Wolfram was excited about Mistral’s return to Apache 2.0, Yam couldn’t stop talking about Claude Opus 4.5 after a full week of using it, and Nisten came out of left field with an AWQ quantization of Prime Intellect’s model that apparently runs incredibly fast on a single GPU. As for me? I’m torn between Opus 4.5 (which literally fixed bugs that Gemini 3 created in my code) and DeepSeek’s gold-medal winning reasoning model.

Speaking of which, let’s dive into what happened this week, starting with the open source stuff that’s been absolutely cooking.

Open Source LLMs

DeepSeek V3.2: The Whale Returns with Gold Medals

The whale is back, folks! DeepSeek released two major updates this week: V3.2 and V3.2-Speciale. And these aren’t incremental improvements—we’re talking about an open reasoning-first model that’s rivaling GPT-5 and Gemini 3 Pro with actual gold medal Olympiad wins.

Here’s what makes this release absolutely wild: DeepSeek V3.2-Speciale is achieving 96% on AIME versus 94% for GPT-5 High. It’s getting gold medals on IMO (35/42), CMO, ICPC (10/12), and IOI (492/600). This is a 685 billion parameter MOE model with MIT license, and it literally broke the benchmark graph on HMMT 2025—the score was so high it went outside the chart boundaries. That’s how you DeepSeek, basically.

But it’s not just about reasoning. The regular V3.2 (not Speciale) is absolutely crushing it on agentic benchmarks: 73.1% on SWE-Bench Verified, first open model over 35% on Tool Decathlon, and 80.3% on τ²-bench. It’s now the second most intelligent open weights model and ranks ahead of Grok 4 and Claude Sonnet 4.5 on Artificial Analysis.

The price is what really makes this insane: 28 cents per million tokens on OpenRouter. That’s absolutely ridiculous for this level of performance. They’ve also introduced DeepSeek Sparse Attention (DSA) which gives you 2-3x cheaper 128K inference without performance loss. LDJ pointed out on the show that he appreciates how transparent they’re being about not quite matching Gemini 3’s efficiency on reasoning tokens, but it’s open source and incredibly cheap.

One thing to note: V3.2-Speciale doesn’t support tool calling. As Wolfram pointed out from the model card, it’s “designed exclusively for deep reasoning tasks.” So if you need agentic capabilities, stick with the regular V3.2.

Check out the full release on Hugging Face or read the announcement.

Mistral 3: Europe’s Favorite AI Lab Returns to Apache 2.0

Mistral is back, and they’re back with fully open Apache 2.0 licenses across the board! This is huge news for the open source community. They released two major things this week: Mistral Large 3 and the Ministral 3 family of small models.

Mistral Large 3 is a 675 billion parameter MOE with 41 billion active parameters and a quarter million (256K) context window, trained on 3,000 H200 GPUs. There’s been some debate about this model’s performance, and I want to address the elephant in the room: some folks saw a screenshot showing Mistral Large 3 very far down on Artificial Analysis and started dunking on it. But here’s the key context that Merve from Hugging Face pointed out—this is the only non-reasoning model on that chart besides GPT 5.1. When you compare it to other instruction-tuned (non-reasoning) models, it’s actually performing quite well, sitting at #6 among open models on LMSys Arena.

Nisten checked LM Arena and confirmed that on coding specifically, Mistral Large 3 is scoring as one of the best open source coding models available. Yam made an important point that we should compare Mistral to other open source players like Qwen and DeepSeek rather than to closed models—and in that context, this is a solid release.

But the real stars of this release are the Ministral 3 small models: 3B, 8B, and 14B, all with vision capabilities. These are edge-optimized, multimodal, and the 3B actually runs completely in the browser with WebGPU using transformers.js. The 14B reasoning variant achieves 85% on AIME 2025, which is state-of-the-art for its size class. Wolfram confirmed that the multilingual performance is excellent, particularly for German.

There’s been some discussion about whether Mistral Large 3 is a DeepSeek finetune given the architectural similarities, but Mistral claims these are fully trained models. As Nisten noted, even if they used similar architecture (which is Apache 2.0 licensed), there’s nothing wrong with that—it’s an excellent architecture that works. Lucas Atkins later confirmed on the show that “Mistral Large looks fantastic... it is DeepSeek through and through architecture wise. But Kimi also does that—DeepSeek is the GOAT. Training MOEs is not as easy as just import deepseak and train.”

Check out Mistral Large 3 and Ministral 3 on Hugging Face.

Arcee Trinity: US-Trained MOEs Are Back

We had Lucas Atkins, CTO of Arcee AI, join us on the show to talk about their new Trinity family of models, and this conversation was packed with insights about what it takes to train MOEs from scratch in the US.

Trinity is a family of open-weight MOEs fully trained end-to-end on American infrastructure with 10 trillion curated tokens from Datology.ai. They released Trinity-Mini (26B total, 3B active) and Trinity-Nano-Preview (6B total, 1B active), with Trinity-Large (420B parameters, 13B active) coming in mid-January 2026.

The benchmarks are impressive: Trinity-Mini hits 84.95% on MMLU (0-shot), 92.1% on Math-500, and 65% on GPQA Diamond. But what really caught my attention was the inference speed—Nano generates at 143 tokens per second on llama.cpp, and Mini hits 157 t/s on consumer GPUs. They’ve even demonstrated it running on an iPhone via MLX Swift.

I asked Lucas why it matters where models come from, and his answer was nuanced: for individual developers, it doesn’t really matter—use the best model for your task. But for Fortune 500 companies, compliance and legal teams are getting increasingly particular about where models were trained and hosted. This is slowing down enterprise AI adoption, and Trinity aims to solve that.

Lucas shared a fascinating insight about why they decided to do full pretraining instead of just post-training on other people’s checkpoints: “We at Arcee were relying on other companies releasing capable open weight models... I didn’t like the idea of the foundation of our business being reliant on another company releasing models.” He also dropped some alpha about Trinity-Large: they’re going with 13B active parameters instead of 32B because going sparser actually gave them much faster throughput on Blackwell GPUs.

The conversation about MOEs being cheaper for RL was particularly interesting. Lucas explained that because MOEs are so inference-efficient, you can do way more rollouts during reinforcement learning, which means more RL benefit per compute dollar. This is likely why we’re seeing labs like MiniMax go from their original 456B/45B-active model to a leaner 220B/10B-active model—they can get more gains in post-training by being able to do more steps.

Check out Trinity-Mini and Trinity-Nano-Preview on Hugging Face, or read The Trinity Manifesto.

OpenAI Code Red: Panic at the Disco (and Garlic?)

It was ChatGPT’s 3rd birthday this week (Nov 30th), but the party vibes seem… stressful. Reports came out that Sam Altman has declared a “Code Red” at OpenAI.

Why? Gemini 3.The user numbers don’t lie. ChatGPT apparently saw a 6% drop in daily active users following the Gemini 3 launch. Google’s integration is just too good, and their free tier is compelling.

In response, OpenAI has supposedly paused “side projects” (ads, shopping bots) to focus…