I just saw the Kimi K2 Thinking release!
Kimi K2 is based on the DeepSeek V3/R1 architecture, and here's a side-by-side comparison.
In short, Kimi K2 is a slightly scaled DeepSeek V3/R1. And the gains are in the data and training recipes. Hopefully, we will see some details on those soon, too.
Kimi K2 is based on the DeepSeek V3/R1 architecture, and here's a side-by-side comparison.
In short, Kimi K2 is a slightly scaled DeepSeek V3/R1. And the gains are in the data and training recipes. Hopefully, we will see some details on those soon, too.
November 6, 2025 at 7:35 PM
I just saw the Kimi K2 Thinking release!
Kimi K2 is based on the DeepSeek V3/R1 architecture, and here's a side-by-side comparison.
In short, Kimi K2 is a slightly scaled DeepSeek V3/R1. And the gains are in the data and training recipes. Hopefully, we will see some details on those soon, too.
Kimi K2 is based on the DeepSeek V3/R1 architecture, and here's a side-by-side comparison.
In short, Kimi K2 is a slightly scaled DeepSeek V3/R1. And the gains are in the data and training recipes. Hopefully, we will see some details on those soon, too.
K2-Thinking is SOTA, top model in agentic tool calling
November 7, 2025 at 10:40 AM
K2-Thinking is SOTA, top model in agentic tool calling
Best breakdown of modern LLM architectures
From DeepSeek to GPT-OSS, it’s all here ↓
Covers every flagship model
1️⃣ DeepSeek V3/R1
2️⃣ OLMo 2
3️⃣ Gemma 3
4️⃣ Mistral Small 3.1
5️⃣ Llama 4
6️⃣ Qwen3
7️⃣ SmolLM3
8️⃣ Kimi 2
9️⃣ GPT-OSS
#ArtificialIntelligence #MachineLearning #DeepLearning #DataScience #Analytics
From DeepSeek to GPT-OSS, it’s all here ↓
Covers every flagship model
1️⃣ DeepSeek V3/R1
2️⃣ OLMo 2
3️⃣ Gemma 3
4️⃣ Mistral Small 3.1
5️⃣ Llama 4
6️⃣ Qwen3
7️⃣ SmolLM3
8️⃣ Kimi 2
9️⃣ GPT-OSS
#ArtificialIntelligence #MachineLearning #DeepLearning #DataScience #Analytics
November 7, 2025 at 12:27 PM
Best breakdown of modern LLM architectures
From DeepSeek to GPT-OSS, it’s all here ↓
Covers every flagship model
1️⃣ DeepSeek V3/R1
2️⃣ OLMo 2
3️⃣ Gemma 3
4️⃣ Mistral Small 3.1
5️⃣ Llama 4
6️⃣ Qwen3
7️⃣ SmolLM3
8️⃣ Kimi 2
9️⃣ GPT-OSS
#ArtificialIntelligence #MachineLearning #DeepLearning #DataScience #Analytics
From DeepSeek to GPT-OSS, it’s all here ↓
Covers every flagship model
1️⃣ DeepSeek V3/R1
2️⃣ OLMo 2
3️⃣ Gemma 3
4️⃣ Mistral Small 3.1
5️⃣ Llama 4
6️⃣ Qwen3
7️⃣ SmolLM3
8️⃣ Kimi 2
9️⃣ GPT-OSS
#ArtificialIntelligence #MachineLearning #DeepLearning #DataScience #Analytics
this morning, X is saturated with people from US claiming that their favorite unknown benchmark (that happens to show K2 trailing US models) is actually the best single benchmark to watch
lol notice how they clipped off the top 12
lol notice how they clipped off the top 12
November 8, 2025 at 12:10 PM
this morning, X is saturated with people from US claiming that their favorite unknown benchmark (that happens to show K2 trailing US models) is actually the best single benchmark to watch
lol notice how they clipped off the top 12
lol notice how they clipped off the top 12
the funniest post from around the deepseek r1 release
October 30, 2025 at 4:50 PM
the funniest post from around the deepseek r1 release
Running #GPU workloads on #Kubernetes with #TalosLinux isn’t like using traditional Linux.
Here's how to deploy the Deepseek-r1 LLM using Ollama on bare metal Kubernetes with Talos and Omni’s Image Factory. → www.youtube.com/watch?v=HiDW...
Want to talk more about it? Find our team at #KubeCon!
Here's how to deploy the Deepseek-r1 LLM using Ollama on bare metal Kubernetes with Talos and Omni’s Image Factory. → www.youtube.com/watch?v=HiDW...
Want to talk more about it? Find our team at #KubeCon!
Deepseek on bare metal Kubernetes with Talos Linux
Starting from a blank computer with an NVIDIA GPU we walk through all the steps needed to deploy Deepseek-r1 as a Kubernetes workload.
Sign up for Omni at https://siderolabs.com/omni-signup
www.youtube.com
October 28, 2025 at 5:06 PM
Running #GPU workloads on #Kubernetes with #TalosLinux isn’t like using traditional Linux.
Here's how to deploy the Deepseek-r1 LLM using Ollama on bare metal Kubernetes with Talos and Omni’s Image Factory. → www.youtube.com/watch?v=HiDW...
Want to talk more about it? Find our team at #KubeCon!
Here's how to deploy the Deepseek-r1 LLM using Ollama on bare metal Kubernetes with Talos and Omni’s Image Factory. → www.youtube.com/watch?v=HiDW...
Want to talk more about it? Find our team at #KubeCon!
New post recapping the biggest paper in reasoning/RL this year since the DeepSeek R1 report. It does a great job highlighting how RL (and post training really) is far more of an art than a science (pretraining is a hard science).
www.interconnects.ai/p/the-new-rl...
www.interconnects.ai/p/the-new-rl...
The new RL scaling laws
The most covetable research.
www.interconnects.ai
October 20, 2025 at 3:28 PM
New post recapping the biggest paper in reasoning/RL this year since the DeepSeek R1 report. It does a great job highlighting how RL (and post training really) is far more of an art than a science (pretraining is a hard science).
www.interconnects.ai/p/the-new-rl...
www.interconnects.ai/p/the-new-rl...
Agents are hard to benchmark
new research from Princeton shows several factors that complect benchmarking
agents will:
- take shortcuts
- take overly expensive actions
- hardcode answers
also, token efficiency doesn’t translate to cost reduction
arxiv.org/abs/2510.11977
new research from Princeton shows several factors that complect benchmarking
agents will:
- take shortcuts
- take overly expensive actions
- hardcode answers
also, token efficiency doesn’t translate to cost reduction
arxiv.org/abs/2510.11977
October 16, 2025 at 11:51 AM
Agents are hard to benchmark
new research from Princeton shows several factors that complect benchmarking
agents will:
- take shortcuts
- take overly expensive actions
- hardcode answers
also, token efficiency doesn’t translate to cost reduction
arxiv.org/abs/2510.11977
new research from Princeton shows several factors that complect benchmarking
agents will:
- take shortcuts
- take overly expensive actions
- hardcode answers
also, token efficiency doesn’t translate to cost reduction
arxiv.org/abs/2510.11977
Aaahhh I might have to do some painfully slow inference so I can see the CoT from r1-zero. I want my models untamed, SFT-free, speaking in tongues.
huggingface.co/deepseek-ai/...
huggingface.co/deepseek-ai/...
deepseek-ai/DeepSeek-R1-Zero · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co
October 16, 2025 at 4:45 AM
Aaahhh I might have to do some painfully slow inference so I can see the CoT from r1-zero. I want my models untamed, SFT-free, speaking in tongues.
huggingface.co/deepseek-ai/...
huggingface.co/deepseek-ai/...
Chinese LLMs like DeepSeek have been getting good for a while.
See our tracker geni.us/IIB-LLMs
made with @vizsweet
See our tracker geni.us/IIB-LLMs
made with @vizsweet
October 15, 2025 at 8:55 PM
Chinese LLMs like DeepSeek have been getting good for a while.
See our tracker geni.us/IIB-LLMs
made with @vizsweet
See our tracker geni.us/IIB-LLMs
made with @vizsweet
TRM's performance is 🔥! It beat DeepSeek R1 (671B params) & Gemini 2.5 Pro on ARC-AGI benchmark. Achieved 44.6% on ARC-AGI-1 & 87% on Sudoku-Extreme! 💯🏆 #AIbenchmark #DeepLearning
October 14, 2025 at 8:26 AM
TRM's performance is 🔥! It beat DeepSeek R1 (671B params) & Gemini 2.5 Pro on ARC-AGI benchmark. Achieved 44.6% on ARC-AGI-1 & 87% on Sudoku-Extreme! 💯🏆 #AIbenchmark #DeepLearning
deepseek-r1-052b-qwen-8b
35 tok/s
qwen-coder-30b
59 tok/s
gemma-3n-e4b
42 tok/s
gpt-oss-20b
57 tok/s
gpt-oss-120b
27 tok/s
35 tok/s
qwen-coder-30b
59 tok/s
gemma-3n-e4b
42 tok/s
gpt-oss-20b
57 tok/s
gpt-oss-120b
27 tok/s
October 10, 2025 at 9:09 PM
deepseek-r1-052b-qwen-8b
35 tok/s
qwen-coder-30b
59 tok/s
gemma-3n-e4b
42 tok/s
gpt-oss-20b
57 tok/s
gpt-oss-120b
27 tok/s
35 tok/s
qwen-coder-30b
59 tok/s
gemma-3n-e4b
42 tok/s
gpt-oss-20b
57 tok/s
gpt-oss-120b
27 tok/s
Updated & turned my Big LLM Architecture Comparison article into a video lecture.
The 11 LLM archs covered in this video:
1. DeepSeek V3/R1
2. OLMo 2
3. Gemma 3
4. Mistral Small 3.1
5. Llama 4
6. Qwen3
7. SmolLM3
8. Kimi 2
9. GPT-OSS
10. Grok 2.5
11. GLM-4.5/4.6
www.youtube.com/watch?v=rNlU...
The 11 LLM archs covered in this video:
1. DeepSeek V3/R1
2. OLMo 2
3. Gemma 3
4. Mistral Small 3.1
5. Llama 4
6. Qwen3
7. SmolLM3
8. Kimi 2
9. GPT-OSS
10. Grok 2.5
11. GLM-4.5/4.6
www.youtube.com/watch?v=rNlU...
The Big LLM Architecture Comparison
YouTube video by Sebastian Raschka
www.youtube.com
October 10, 2025 at 5:05 PM
Updated & turned my Big LLM Architecture Comparison article into a video lecture.
The 11 LLM archs covered in this video:
1. DeepSeek V3/R1
2. OLMo 2
3. Gemma 3
4. Mistral Small 3.1
5. Llama 4
6. Qwen3
7. SmolLM3
8. Kimi 2
9. GPT-OSS
10. Grok 2.5
11. GLM-4.5/4.6
www.youtube.com/watch?v=rNlU...
The 11 LLM archs covered in this video:
1. DeepSeek V3/R1
2. OLMo 2
3. Gemma 3
4. Mistral Small 3.1
5. Llama 4
6. Qwen3
7. SmolLM3
8. Kimi 2
9. GPT-OSS
10. Grok 2.5
11. GLM-4.5/4.6
www.youtube.com/watch?v=rNlU...
why would DeepSeek drop the R1 brand and not name the next model “R2”?
i get that people in AI are bad at branding, but are they really *this* bad?
afaict the next one is V4, but they got so much publicity with R1..
i get that people in AI are bad at branding, but are they really *this* bad?
afaict the next one is V4, but they got so much publicity with R1..
October 2, 2025 at 9:32 PM
why would DeepSeek drop the R1 brand and not name the next model “R2”?
i get that people in AI are bad at branding, but are they really *this* bad?
afaict the next one is V4, but they got so much publicity with R1..
i get that people in AI are bad at branding, but are they really *this* bad?
afaict the next one is V4, but they got so much publicity with R1..
Researchers assessed 30 game concepts with midsize LLMs—LLaMA 3.1, Qwen 2.5 and DeepSeek‑R1—and found DeepSeek‑R1 gave useful feedback. The rubric covered narrative hook, mechanics and market potential. https://getnews.me/medium-sized-llms-show-promise-for-early-game-design-feedback/ #llm #gamedev
September 30, 2025 at 11:28 PM
Researchers assessed 30 game concepts with midsize LLMs—LLaMA 3.1, Qwen 2.5 and DeepSeek‑R1—and found DeepSeek‑R1 gave useful feedback. The rubric covered narrative hook, mechanics and market potential. https://getnews.me/medium-sized-llms-show-promise-for-early-game-design-feedback/ #llm #gamedev
The paper says they bake in a system prompt as part of the RL process:
"2.2.3. Training Template
To train DeepSeek-R1-Zero, we begin by designing a straightforward template that guides the base model to adhere to our specified instructions."
"2.2.3. Training Template
To train DeepSeek-R1-Zero, we begin by designing a straightforward template that guides the base model to adhere to our specified instructions."
September 29, 2025 at 1:00 PM
The paper says they bake in a system prompt as part of the RL process:
"2.2.3. Training Template
To train DeepSeek-R1-Zero, we begin by designing a straightforward template that guides the base model to adhere to our specified instructions."
"2.2.3. Training Template
To train DeepSeek-R1-Zero, we begin by designing a straightforward template that guides the base model to adhere to our specified instructions."
Anthropic's Claude 3.7 Sonnet is the new king 👑 of code generation (but only with help), and DeepSeek R1 disappoints buff.ly/BUvGPLL
#Java #CodeGen #genai #llm
#Java #CodeGen #genai #llm
September 24, 2025 at 5:08 AM
Anthropic's Claude 3.7 Sonnet is the new king 👑 of code generation (but only with help), and DeepSeek R1 disappoints buff.ly/BUvGPLL
#Java #CodeGen #genai #llm
#Java #CodeGen #genai #llm
I feel very proud to be part of @nature.com, and to have colleagues who handled this excellent #DeepSeek paper that describes DeepSeek-R1, because it's the first widely used commercial LLM that has been published in a peer-reviewed journal 🧪 www.nature.com/articles/s41...
DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning - Nature
A new artificial intelligence model, DeepSeek-R1, is introduced, demonstrating that the reasoning abilities of large language models can be incentivized through pure reinforcement learning, removing t...
www.nature.com
September 22, 2025 at 6:12 PM
I feel very proud to be part of @nature.com, and to have colleagues who handled this excellent #DeepSeek paper that describes DeepSeek-R1, because it's the first widely used commercial LLM that has been published in a peer-reviewed journal 🧪 www.nature.com/articles/s41...
Huawei says DeepSeek-R1-Safe, which was trained on 1,000 of its Ascend AI chips, is "nearly 100% successful" in preventing politically sensitive topics (Eduardo Baptista/Reuters)
Main Link | Techmeme Permalink
Main Link | Techmeme Permalink
September 20, 2025 at 1:41 PM
Huawei says DeepSeek-R1-Safe, which was trained on 1,000 of its Ascend AI chips, is "nearly 100% successful" in preventing politically sensitive topics (Eduardo Baptista/Reuters)
Main Link | Techmeme Permalink
Main Link | Techmeme Permalink
nella tarda serata di giovedì di aver utilizzato 1.000 dei suoi chip Ascend AI per addestrare il modello a linguaggio esteso, che è stato ottimizzato dal modello open source R1 di DeepSeek
Il partner di Huawei era l'élite della Zhejiang University
l'alma mater del fondatore di DeepSeek ⬇️
Il partner di Huawei era l'élite della Zhejiang University
l'alma mater del fondatore di DeepSeek ⬇️
September 20, 2025 at 10:34 AM
nella tarda serata di giovedì di aver utilizzato 1.000 dei suoi chip Ascend AI per addestrare il modello a linguaggio esteso, che è stato ottimizzato dal modello open source R1 di DeepSeek
Il partner di Huawei era l'élite della Zhejiang University
l'alma mater del fondatore di DeepSeek ⬇️
Il partner di Huawei era l'élite della Zhejiang University
l'alma mater del fondatore di DeepSeek ⬇️
"DeepSeek didn’t really train its flagship model for $294,000: Training costs detailed in R1 training report don't include 2.79 million GPU hours that laid its foundation"
Counterpoint: It cost a lot more than $294k.
🤷♂️
www.theregister.com/2025/09/19/d...
Counterpoint: It cost a lot more than $294k.
🤷♂️
www.theregister.com/2025/09/19/d...
DeepSeek didn’t really train its flagship model for $294,000
: Training costs detailed in R1 training report don't include 2.79 million GPU hours that laid its foundation
www.theregister.com
September 19, 2025 at 6:12 PM
"DeepSeek didn’t really train its flagship model for $294,000: Training costs detailed in R1 training report don't include 2.79 million GPU hours that laid its foundation"
Counterpoint: It cost a lot more than $294k.
🤷♂️
www.theregister.com/2025/09/19/d...
Counterpoint: It cost a lot more than $294k.
🤷♂️
www.theregister.com/2025/09/19/d...
An editorial was published in Nature recently claiming that glam journal publication of LLMs (like DeepSeek-R1 this case) marks a step towards greater transparency, accountability & credibility www.nature.com/articles/d41.... I have thoughts ... 1/
https://www.nature.com/articles/d41586-025-02979-9
t.co
September 19, 2025 at 7:01 PM
An editorial was published in Nature recently claiming that glam journal publication of LLMs (like DeepSeek-R1 this case) marks a step towards greater transparency, accountability & credibility www.nature.com/articles/d41.... I have thoughts ... 1/
Nature research paper: DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning
go.nature.com/41WGjPu
go.nature.com/41WGjPu
DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning - Nature
A new artificial intelligence model, DeepSeek-R1, is introduced, demonstrating that the reasoning abilities of large language models can be incentivized through pure reinforcement learning, removing the need for human-annotated demonstrations.
go.nature.com
September 19, 2025 at 8:46 AM
Nature research paper: DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning
go.nature.com/41WGjPu
go.nature.com/41WGjPu