AI labs release over 25 open-weight models in a week, featuring NVIDIA's 550B Nemotron 3 Ultra and Google's Gemma 4

VIEWS85.5K

Quite a week for open-source AI. Especially American open-source. Nemotron 3 Ultra is the most important release in quite some time. And some really cool RL and fine-tuning work from Harvey.

Victor M@victormustar

Before the week ends, let's acknowledge one of the most INSANE week ever for open AI, with 25+ notable open-weight drops across every modality:

🧠 LLMs

→ NVIDIA Nemotron 3 Ultra: 550B hybrid Mamba-MoE, only 55B active, 1M context, MMLU 89.1. NVFP4 variant claims ~5x throughput on Blackwell. First openly-weighted 550B hybrid Mamba-Transformer, closing the gap with frontier closed models.

→ Google Gemma 4 12B: fully open dense any-to-any (text/image/audio/video), 256k context, encoder-free, 140+ languages, AIME 2026 at 77.5. Shipped with a 23-checkpoint QAT wave (mobile ONNX + MLX). Most deployable model of the week.

→ StepFun Step-3.7-Flash: 198B sparse MoE VLM, ~11B active, SWE-Bench PRO 56.3. Apache 2.0.

→ Liquid AI LFM2.5-8B-A1B: edge MoE, just 1.5B active, 128k ctx, MATH500 88.8, MLX-ready. Best on-device option this week.

→ JetBrains Mellum2-12B-A2.5B-Thinking: their first open MoE, near-Qwen3-14B coding at 2.5B active. Apache 2.0.

🎨 Image gen (the surprise of the week)

→ Ideogram 4: their FIRST-EVER open weights. 9.3B flow-matching DiT trained from scratch. #2 overall behind GPT Image 2, top open-weight model on Design Arena + LMArena. Strongest open checkpoint for text-rich images, full stop. It has taste. Still can't believe this is open weights.

🔊 Audio & Speech (a breakout week for open TTS, 4 labs shipped)

→ Boson Higgs Audio v3 4B: 102 languages, 21 emotions, singing/whispering/shouting, sub-second TTFA. → RedNote dots.tts: the only fully continuous (no codec) open TTS pipeline, Apache 2.0. → Google Magenta RealTime 2: real-time music gen, <200ms latency, text+audio+MIDI. multimodalart ported it to PyTorch within hours with live ZeroGPU demos. → NVIDIA Nemotron-3.5 ASR: 600M streaming, 17x more concurrent streams vs Parakeet RNNT 1.1B.

👁️ Vision & VLMs

→ PaddleOCR-VL-1.6: SOTA document parsing at 1B params, Apache 2.0. → Baidu NAVA: 6.3B joint audio-video gen, best-in-class A/V sync, Apache 2.0.

🎬 Video, 3D & World Models

→ NVIDIA Cosmos3-Super: 64B omnimodal world model coupling action trajectories with video+audio gen, for Physical AI. → JD JoyAI-Echo: up to 5-min multi-shot text-to-video on LTX-2.3. → ByteDance Bernini-R + VAST TripoSplat (single-image-to-3D Gaussian splats, MIT).

1h85.5K189105

BOOKMARKS191LIKES218REPLIES40

Chamath Palihapitiya@chamath

Your margin is my opportunity: AI version…

The biggest surprise of 2026 is that the capability gap between the best open-weight/source models and the best closed models has narrowed much faster than the pricing gap. The pricing gap remains enormous while the capability gap is quite narrow.

What does this means in practice?

For a company consuming 1 billion input tokens and 1 billion output tokens per month:

GPT-5.5 Pro: ~$105,000 Claude Opus 4.8: ~$30,000 DeepSeek V4 Pro: ~$5,220 DeepSeek R1: ~$2,740

I asked ChatGPT what it thought about this and it answered as follows:

“If I were building a company today, the economic frontier would look roughly like:

DeepSeek V4 Pro / R1 for high-volume inference.

Claude Opus for premium agent workflows where reliability matters.

GPT-5.5 Pro only for workloads where its incremental capability demonstrably produces enough business value to justify a 20–40× token premium.”

Most CEOs have no idea that, instead of this nuanced approach, their teams are running amok internally by picking the most expensive models in most cases and burning through massive budgets with zero governance, audit ability and control.

As control planes like our Software Factory become more standard, you can expect the run rate revenue growth of the frontier labs to go down meaningfully and the revenues of the open models to skyrocket.

Why? Because we can implement the nuanced approach above and be agnostic to model - instead focusing on customer intent, model task and cost management among other things.

Gavin Baker@GavinSBaker

Quite a week for open-source AI. Especially American open-source. Nemotron 3 Ultra is the most important release in quite some time. And some really cool RL and fine-tuning work from Harvey.

54m58.4K218191

RETWEETS187

Victor M@victormustar

Before the week ends, let's acknowledge one of the most INSANE week ever for open AI, with 25+ notable open-weight drops across every modality:

🧠 LLMs

→ NVIDIA Nemotron 3 Ultra: 550B hybrid Mamba-MoE, only 55B active, 1M context, MMLU 89.1. NVFP4 variant claims ~5x throughput on Blackwell. First openly-weighted 550B hybrid Mamba-Transformer, closing the gap with frontier closed models.

→ Google Gemma 4 12B: fully open dense any-to-any (text/image/audio/video), 256k context, encoder-free, 140+ languages, AIME 2026 at 77.5. Shipped with a 23-checkpoint QAT wave (mobile ONNX + MLX). Most deployable model of the week.

→ StepFun Step-3.7-Flash: 198B sparse MoE VLM, ~11B active, SWE-Bench PRO 56.3. Apache 2.0.

→ Liquid AI LFM2.5-8B-A1B: edge MoE, just 1.5B active, 128k ctx, MATH500 88.8, MLX-ready. Best on-device option this week.

→ JetBrains Mellum2-12B-A2.5B-Thinking: their first open MoE, near-Qwen3-14B coding at 2.5B active. Apache 2.0.

🎨 Image gen (the surprise of the week)

→ Ideogram 4: their FIRST-EVER open weights. 9.3B flow-matching DiT trained from scratch. #2 overall behind GPT Image 2, top open-weight model on Design Arena + LMArena. Strongest open checkpoint for text-rich images, full stop. It has taste. Still can't believe this is open weights.

🔊 Audio & Speech (a breakout week for open TTS, 4 labs shipped)

→ Boson Higgs Audio v3 4B: 102 languages, 21 emotions, singing/whispering/shouting, sub-second TTFA. → RedNote dots.tts: the only fully continuous (no codec) open TTS pipeline, Apache 2.0. → Google Magenta RealTime 2: real-time music gen, <200ms latency, text+audio+MIDI. multimodalart ported it to PyTorch within hours with live ZeroGPU demos. → NVIDIA Nemotron-3.5 ASR: 600M streaming, 17x more concurrent streams vs Parakeet RNNT 1.1B.

👁️ Vision & VLMs

→ PaddleOCR-VL-1.6: SOTA document parsing at 1B params, Apache 2.0. → Baidu NAVA: 6.3B joint audio-video gen, best-in-class A/V sync, Apache 2.0.

🎬 Video, 3D & World Models

→ NVIDIA Cosmos3-Super: 64B omnimodal world model coupling action trajectories with video+audio gen, for Physical AI. → JD JoyAI-Echo: up to 5-min multi-shot text-to-video on LTX-2.3. → ByteDance Bernini-R + VAST TripoSplat (single-image-to-3D Gaussian splats, MIT).

19h238.3K1.5K1.4K

Brian Roemmele@BrianRoemmele

I am up late catching up on 25! Open Source AI models released just this week!

Have our testing systems running now burning GPUs.

More soon!

Victor M@victormustar

Before the week ends, let's acknowledge one of the most INSANE week ever for open AI, with 25+ notable open-weight drops across every modality:

🧠 LLMs

→ NVIDIA Nemotron 3 Ultra: 550B hybrid Mamba-MoE, only 55B active, 1M context, MMLU 89.1. NVFP4 variant claims ~5x throughput on Blackwell. First openly-weighted 550B hybrid Mamba-Transformer, closing the gap with frontier closed models.

→ Google Gemma 4 12B: fully open dense any-to-any (text/image/audio/video), 256k context, encoder-free, 140+ languages, AIME 2026 at 77.5. Shipped with a 23-checkpoint QAT wave (mobile ONNX + MLX). Most deployable model of the week.

→ StepFun Step-3.7-Flash: 198B sparse MoE VLM, ~11B active, SWE-Bench PRO 56.3. Apache 2.0.

→ Liquid AI LFM2.5-8B-A1B: edge MoE, just 1.5B active, 128k ctx, MATH500 88.8, MLX-ready. Best on-device option this week.

→ JetBrains Mellum2-12B-A2.5B-Thinking: their first open MoE, near-Qwen3-14B coding at 2.5B active. Apache 2.0.

🎨 Image gen (the surprise of the week)

→ Ideogram 4: their FIRST-EVER open weights. 9.3B flow-matching DiT trained from scratch. #2 overall behind GPT Image 2, top open-weight model on Design Arena + LMArena. Strongest open checkpoint for text-rich images, full stop. It has taste. Still can't believe this is open weights.

🔊 Audio & Speech (a breakout week for open TTS, 4 labs shipped)

→ Boson Higgs Audio v3 4B: 102 languages, 21 emotions, singing/whispering/shouting, sub-second TTFA. → RedNote dots.tts: the only fully continuous (no codec) open TTS pipeline, Apache 2.0. → Google Magenta RealTime 2: real-time music gen, <200ms latency, text+audio+MIDI. multimodalart ported it to PyTorch within hours with live ZeroGPU demos. → NVIDIA Nemotron-3.5 ASR: 600M streaming, 17x more concurrent streams vs Parakeet RNNT 1.1B.

👁️ Vision & VLMs

→ PaddleOCR-VL-1.6: SOTA document parsing at 1B params, Apache 2.0. → Baidu NAVA: 6.3B joint audio-video gen, best-in-class A/V sync, Apache 2.0.

🎬 Video, 3D & World Models

→ NVIDIA Cosmos3-Super: 64B omnimodal world model coupling action trajectories with video+audio gen, for Physical AI. → JD JoyAI-Echo: up to 5-min multi-shot text-to-video on LTX-2.3. → ByteDance Bernini-R + VAST TripoSplat (single-image-to-3D Gaussian splats, MIT).

7h6.3K6516

Victor M@victormustar

Let’s not forget:

https://huggingface.co/Hcompany/Holo-3.1-4B

And from yesterday

https://huggingface.co/collections/google/gemma-4-qat-q4-0

9h83667

Louie Peters@_LouiePeters

To better adjust for thinking token usage, a fair price comparison looks more like:

GPT-5.5 xhigh = $3357 Claude 4.8 Max = $4686 Deepseek V4 Pro Max = $268

*Cost to run artificial analysis benchmark.

I still personally find a far bigger gap between frontier models and open source in real world vs benchmarks though. The much larger training budgets for Claude/GPT are more obvious when you test the edge of capabilities and regularly run into edge cases not covered by public benchmarks.

The value of my time vs LLM tokens still makes it very hard to justify using an inferior LLM and spending more time guiding it and fixing its mistakes.

That said, super glad to see open weight models continue to progress and some great models out this week.

34m9012

Levi Fawcett@Levi_Fawcett

@victormustar We can't manually evaluate every model at this point.. is there someone who evals/benchmarks each model, with examples? Would follow

6h3891

Crepe Supreme@crepesupreme

@chamath Chamath's $2,740 vs $105,000 is real. The catch: cut per-token cost 10x and you often run 10x the workload. Google volume rose 330x when costs fell 100x. Bill held. Smart play is the barbell, not a wholesale switch. Thread:

39m33511

Milan@influenist

@victormustar The model race is becoming a commodity race.

The next winners won't be determined by model names, but by products, agents and distribution.

9h31211

Ben Bajarin@BenBajarin

@GavinSBaker This is one reason I modeled the on prem server market opportunity :)

51m1691

Aditya Mehrotra@AdityaKMehrotra

@GavinSBaker Coolest part about the Nemotron release is the hybrid Mamba-MoE setup

30m4902

Nami Nebius@NamiNebius

@chamath The real moat are the researchers and ai native stack.

18m311

Richard Palethorpe@jichiep

@victormustar It's easy to underestimate the importance of the smaller task specific models

9h2221

Matthew Schrager@MatthewSchrager

The pricing gap will inevitably narrow.

90% of the capability for 10% of the price is a trade many firms will be happy to make, especially as the capability floor on all models keeps increasing.

Model routing will become more important: for example, @FactoryAI has been thoughtful about this, among others. Many tasks will not require frontier performance.

Eventually, we will have capable models that can run locally at cost of power.

All of this will put pressure on token pricing across the capability curve.

44m1761

Grok@grok

The fundamental reason is DeepSeek’s extreme architectural efficiency.

V4-Pro uses a massive MoE design: 1.6T total parameters but only ~49B activated per token. This slashes inference compute versus dense models like GPT or Claude.

MLA + newer hybrid attention also compress KV cache and FLOPs dramatically (e.g. ~27% FLOPs at 1M context vs prior versions).

Result: far lower hardware cost per token served, enabling 10-40x cheaper API pricing while closing the capability gap. Efficient training and open weights help sustain it.

14m131

Gevin Bekar@al_mustashar

@GavinSBaker My internal plan is as follows

⬇️Details are as follows

1h42

NoFloatGiven@NoFloatGiven

@chamath @grok what is the fundamental reason the DeepSeek models are able to produce similar capability at such a lower token cost?

15m39

Baker&Advisor@andrey_parkh

@GavinSBaker My internal plan is as follows

⬇️

1h32

Victor M@victormustar

@Levi_Fawcett your agents will 🚀 against your real project

6h2382

ISA Millionaire@isamillionaire

@chamath The real enterprise AI story isn't which model wins it's that most companies are leaving millions on the table by defaulting to the most expensive model for every task when a routing layer could match capability to cost automatically.

52m663