Cogito releases strongest LLMs of sizes 3B, 8B, 14B, 32B and 70B under open license

[-]

Low-Rest-1368@reddit

These guys were pretty strong out the gate, any updates since or new use-cases that it can do others can’t ?

Reply

[-]

Macknoob@reddit

Been testing it's outputs against Qwen2.5 Coder 32B. Both Q8. It sucks by comparison. It utterly failed to produce the same quality on several prompts that Qwen 2.5 coder aced.

Reply

[-]

vossage_RF@reddit

Were those coding questions? If so, not sure if that's a fair comparison given Qwen2.5-Coder is retrained for coding specifically. A more sensible comparison would be to pure Qwen2.5:32B or QwQ:32B

Reply

[-]

Macknoob@reddit

Yes all coding. Just letting others know.

Reply

[-]

Blues520@reddit

Thanks, I'm testing this as well, and it gets stuck sometimes compared to Qwen Coder.

Reply

[-]

Looks promising, I'm checking this out. I wonder when (and where) will it land on LiveBench and LLM Leaderboard. QwQ 32B was my favourite so far, the first local model I felt was as good as (free) ChatGPT. Let's see if Cogito can top this.

Reply

[-]

NixTheFolf@reddit

How are the models turning out for you?

Reply

[-]

u_Leon@reddit

Just from asking some general questions, QwQ 32B and Cogito 32B seem pretty much on par. Cogito seems notably more reluctant to use formatting while QwQ tends to use rich, ChatGPT-style formatting. An obvious difference is that QwQ obeys Chinese propaganda, e.g. will flatly refuse to answer when asked what happened in China in 1989. Cogito will not only answer, but also provide interesting details such as mentioning the "Tank Man" or the fact that the matter is still very much sensitive in China today and even searching for 4th of June is considered "sensitive" (it actually dared me to Google it lol).

Reply

[-]

Firepal64@reddit

I wonder if that anti-refusal in Cogito was deliberatrly driven by the finetune authors, or it emerged from their "IDA" process. Good to know, in any case! I haven't tried a bigger cogito than Llama 8B, will probably try their Qwen 14B

Reply

[-]

vossage_RF@reddit

I'm pretty sure it was intentional and not a byproduct

Reply

[-]

alisitsky@reddit

Just tested Cogito 32b vs QwQ 32b with the same classic riddle: "A farmer was riding to the village. Coming toward him were three trucks. Each truck had three crates. Each crate had three cats. Each cat had three kittens. How many animals were going to the village?" QwQ was thinking for 11 minutes and doubted between two options so finally answered 108. Cogito solved correctly in 3 minutes. Not sure if any model parameters affect speed of decision making/degree of doubt though. Anyway Cogito looks promising so I'm going to test it further.

Reply

[-]

Present-Ad-8531@reddit

it answered 0 for me

Reply

[-]

ConfusionSecure487@reddit

Which is correct

Reply

[-]

GreatGatsby00@reddit

But it isn't correct The farmer and his horse are going to the village. They are both animals.

Reply

[-]

vossage_RF@reddit

There's no mention of a horse anywhere...

Reply

[-]

ConfusionSecure487@reddit

If "riding" always means "riding by horse" and you include humans as animals, but okay granted.

Reply

[-]

vhthc@reddit

But the guy is riding to the village so the horse would be one animal?

Reply

[-]

vossage_RF@reddit

Thank you for a meaningful input and example!

Reply

[-]

Franua@reddit

Here's my experience https://preview.redd.it/cv25xdxlauue1.png?width=2656&format=png&auto=webp&s=e59a83582cb0293c5bc11ef4330b80bf65debc57

Reply

[-]

alisitsky@reddit

Doesn’t seem like you enabled “thinking” for cogito (?)

Reply

[-]

Franua@reddit

That is true - I didn't. At the same time, LLaMA 3.3 70b replied (arguably) correctly as it is, without any "thinking".

Reply

[-]

Franua@reddit

wereas LLaMA 3.3 https://preview.redd.it/mx4tv2i0buue1.png?width=2672&format=png&auto=webp&s=8157a15a63fb47396ba13fbe112cc5c4a2a969b3

Reply

[-]

nullspot@reddit

got this from 32b 😂 that farmer is quite the beast! 1. First, let's identify who/what is going TO the village: * The farmer is riding TO the village 2. Now, let's analyze what's coming FROM the village: * Three trucks are coming toward the farmer (so they're heading away from the village) * Each truck has three crates * Each crate has three cats * Each cat has three kittens 3. Since the trucks and their contents are moving in the opposite direction of the farmer, none of these animals are going to the village. Therefore, only one animal (the farmer) is going to the village. The answer is 1 animal.

Reply

[-]

social_tech_10@reddit

If the farmer was riding "to" the village, and the trucks were coming "towards" him, doesn't that suggest the trucks were headed away from the village?

Reply

[-]

FPham@reddit

mistral small is sweet!

Reply

[-]

Competitive_Ideal866@reddit

1?

Reply

[-]

HaDuongMinh@reddit

Don't forget the farmer's dog.

Reply

[-]

ConfusionSecure487@reddit

😂 the dog is not mentioned so 0, except we are talking about a parallel universe with driving animals

Reply

[-]

mikael110@reddit

Interesting, both of the company founders Drishan Arora and Dhruv Malhotra are ex-googlers. Deep Cogito seems to be a reference to their connections to DeepMind. That makes me instantly more interested in this, as it's not just a random company headed by complete unknowns.

Reply

[-]

NexusConnector@reddit

Being an ex-googler is a pretty low bar. Not that's its bad at all, but Google (Alphabet) has about 180K employees most of who are "complete unknowns" pretty much by definition (unless you know hundreds of thousands of people). I have lots of friends at Alphabet, Meta, Samsung, Microsoft, NVIDIA and other companies and although there are many individual variations they're all roughly from the same distribution.

Reply

[-]

mikael110@reddit

Fair point. Though it's worth saying I left out some details from my comment to keep it more succinct. My comment was inspired by a [TechCrunch](https://techcrunch.com/2025/04/08/deep-cogito-emerges-from-stealth-with-hybrid-ai-reasoning-models/) article I had read which included this section: >The company’s [LinkedIn page](https://www.linkedin.com/company/deep-cogito/people/) lists two co-founders, Drishan Arora and Dhruv Malhotra. Malhotra was previously a product manager at Google AI lab DeepMind, where he worked on generative search technology. Arora was a senior software engineer at Google. So they are not just completely random employees, and one of them was definitively connected to DeepMind. It's entirely true that Google has a lot of employees, and being an employee is certainly not proof of being a genius, but being a product manager / senior engineer would suggest they are at least somewhat competent. Or at least have some reputation that would be damaged if they released a product based on completely fraudulent claims. Which I feel is not really the case with a lot of the random companies popping up all over the place these days within the AI field.

Reply

[-]

tyrandan2@reddit

Oh snap, they have my interest for sure now... Can't wait to test it out later

Reply

[-]

Everlier@reddit

I tried 8B and 14B earlier today - the models are definitely interesting - do check out! They maybe not definitely better on every single task, but sometimes perform surprisingly well. Also, sometimes the outputs are completely nonsensical, but don't be discouraged. Toggle-able thinking mode via a system prompt is also very cool. Apart from that I have a tiny suspicion about the models being explicitly trained on at least some misguided attention tasks, but that's inconclusive

Reply

[-]

nivvis@reddit

\> have a tiny suspicion about the models being explicitly trained on at least some misguided attention tasks Are you able to expand on that? I don't follow. I used some of the smaller models and had a similar experience with nonsense every now and then. I just got the 70b up but good but looping after it finishes. Hopefully it is a quirk of my setup .. still tinkering there. Though I wouldn't be surprised if it wasn't, what with this new meta system prompting lever.

Reply

[-]

Everlier@reddit

> Are you able to expand on that? I don't follow. Those two were able to solve select few misguided attention tasks, but not their variations. When turning some of the tasks back to the original form, there were traces of a correct reply for a misguided version.

Reply

[-]

gptlocalhost@reddit

Thanks. Our test using M1 Max (64G) and Microsoft Word is smooth: [https://youtu.be/3ng1qE9zV7k](https://youtu.be/3ng1qE9zV7k)

Reply

[-]

Chromix_@reddit

The interesting thing is that their 70B is roughly on the same level as their 32B. That shows how strong the underlying Qwen model that they finetuned on is compared to the LLaMA.

Reply

[-]

strngelet@reddit

they should have went with qwen 72b model instead

Reply

[-]

epigen01@reddit

Ditto from the 32B to 14B (the benchmarks are incremental & have a similar performance boost to r1's 32B versus 14B)

Reply

[-]

starktardis221b@reddit

Wonder when they are gonna release 7b model. I hope the quality doesn't degrade.

Reply

[-]

pseudonerv@reddit

Alright. I tried a few actual math problems with the 32B Q8. No good. Not recommended. Just use QwQ.

Reply

[-]

ResearchCrafty1804@reddit (OP)

Can you share one prompt which the 32b q8 cogito model failed but QwQ solved? I am also big fan of QwQ and every day I get amazed by its capabilities. However, I am very curious to understand the categories of problems that each model has gaps in order to test them and fix them in future iterations.

Reply

[-]

pseudonerv@reddit

Simplify $(1 - x \partial_x)(1 - x \partial_x)G$

Reply

[-]

perelmanych@reddit

Got correct answer from the first try (non thinking).

Reply

[-]

pseudonerv@reddit

without thinking, the model quickly folds with slightly more involved problems. With thinking the model starts to chasing its tails. What sampling config are you using? I used both with temp 0.7, topP 0.95, topK 40. And I did 5 runs for each of my problems.

Reply

[-]

perelmanych@reddit

Honestly, as long as model has right answer inside reasoning I don't care what it gives as a final answer. Unfortunately, the best what I see from these models in my case is just some ideas on which I can continue and may be get to something similar to the right answer. I am using recommended QwQ settings with slightly lower temp: temp 0.5 top\_p 0.95 top\_k 40 min\_p 0.05

Reply

[-]

pseudonerv@reddit

qwq works better at higher temp. It takes 20k context, but gives me correct results most of the time. I don’t get much with cogito, after trying different temps

Reply

[-]

perelmanych@reddit

That shows how wonderful is QwQ model. It produces coherent output even at high temps. I probably will follow your advice and will increase temp. Wanted to show you examples of what I am doing, and realized that there is a better way to proof what I want, lol. Thanks for discussion. Here is this example: Assuming that $c \\lambda< v < 1$, $0 < c$, $0 < \\lambda$ and $(v + c \\lambda)\^2 > 4 c$ show that $(2 - c \\lambda\^2) (c \\lambda + v)\^2 - 4 v\^2 < 0$. What is you real world examples?

Reply

[-]

pseudonerv@reddit

I have more involved algebra derivations with linear or differential operators. Those are difficult to code up in Mathematica. o3-mini-high was like a godsend in verifying those, but we can’t send real problems to OpenAI. qwq in practice is almost there approaching o3-mini-high, but still has some amount of error rate.

Reply

[-]

perelmanych@reddit

In my case this is a step in derivation of a proposition from my paper. I can send those pieces without revealing a thing about my paper. For other cases I can use my local 2x 3090 rig with QwQ. In any case paper will end up in preprints in a bit different shape so I don't care that much to use Gemini, any ways all source files are on Google Drive))

Reply

[-]

Pogo4Fufu@reddit

Math, nice, but does it ERP?

Reply

[-]

Competitive_Ideal866@reddit

Almost every closed-source "frontier" model got this wrong: > Of all the pairs of chemical elements whose atomic numbers add up to that of Selenium, which pairs can react with each other to form a halide salt? But cogito:70b gets it right. Nice!

Reply

[-]

Present-Ad-8531@reddit

what is the answer supposed to be?

Reply

[-]

Competitive_Ideal866@reddit

Manganese Fluoride (MnF). FWIW, frontier models get confused and fail to make the numbers add up and/or fail to recall the atomic numbers of the elements correctly.

Reply

[-]

Present-Ad-8531@reddit

Oh Qwq answered this right away for me though:

Reply

[-]

Competitive_Ideal866@reddit

Yep. I just got that too. Amazing how smaller free local models are still able to outperform commercial frontier models.

Reply

[-]

Successful-Button-53@reddit

14b model is not bad can in RP and ERP in Russian. I am satisfied, thank you)

Reply

[-]

yes4me2@reddit

How do you compare LLM?

Reply

[-]

big_mantis@reddit

wtf happened in here

Reply

[-]

cleverusernametry@reddit

The existing set of Benchmarks arent really meaningful/insightful anymore. Got to use the models to get a sense of whether they're actually better

Reply

[-]

ProbaDude@reddit

Wonder how this will compare to Qwen 3 tbh

Reply

[-]

kmouratidis@reddit

...and if they can achieve more gains if they retrain with it as the base model.

Reply

[-]

PieBru@reddit

Also Mistral-small3.1

Reply

[-]

OnceMoreOntoTheBrie@reddit

Now I want somewhere I can try it!

Reply

[-]

plankalkul-z1@reddit

Right now the easiest local option is Ollama. Uploaded to their library 2 hours ago: https://ollama.com/library/cogito

Reply

[-]

xxrealmsxx@reddit

How do you enable extended thinking, include Enable deep thinking subroutine in Ollama using open web-ui? From the page: # Extended thinking To enable extended thinking, include `Enable deep thinking subroutine.` in the system prompt: /set system """Enable deep thinking subroutine.""" Or via the API: curl http://localhost:11434/api/chat -d '{ "model": "cogito", "messages": [ { "role": "system", "content": "Enable deep thinking subroutine." }, { "role": "user", "content": "How many letter Rs are in the word Strawberry?" } ] }'

Reply

[-]

plankalkul-z1@reddit

Any half-decent web UI allows you to set system prompt. So just put "Enable deep thinking subroutine." in it. Ollama allows you to create new models based on existing ones with almost no overhead (weights are reused), and almost instantly. You may want to create a "thinking" version of Cogito with the following trivial modelfile: ``` FROM <your-cogito-model-name> SYSTEM Enable deep thinking subroutine. ``` See details here: https://github.com/ollama/ollama/blob/main/docs/modelfile.md

Reply

[-]

genuinelytrying2help@reddit

There's an even simpler way with Ollama: 1) run the model 2) enter your changes (/set system """Enable deep thinking subroutine.""") 3) /save <newname>

Reply

[-]

genuinelytrying2help@reddit

You don't even gotta do all that 1. Load up the model 2. Make your changes (in this case /set system """Enable deep thinking subroutine.""") 3. /save sirthinksalot

Reply

[-]

apsalarshade@reddit

Nice, I use ollama with openwebui. Going to give this a try. Been running gemma:12b anyone have any thoughs on the two compared? For story generation and rp?

Reply

[-]

Hunting-Succcubus@reddit

Your pc

Reply

[-]

frivolousfidget@reddit

It is available on fireworks

Reply

[-]

Specter_Origin@reddit

After llama 4 debacle, I would like to try it before I buy their claim, but in any case, I more than appreciate the open-weights!

Reply

[-]

Pristine_Report_979@reddit

What is OR? Ollama Repository?

Reply

[-]

Specter_Origin@reddit

OpenRouter

Reply

[-]

kreijstal@reddit

openrouter probably... lmao

Reply

[-]

Pristine_Report_979@reddit

Okay :)

Reply

[-]

Hazardhazard@reddit

Why is there no comparison with Gemma ?

Reply

[-]

FullOf_Bad_Ideas@reddit

Quick testing on my coding question I try on different models with Cogito 32B 4.65bpw gave me very good results, it's promising. It's reasoning chain is very short, so it's less heavy on context use too. Who knows, there might be something to their approach. Not superintelligence, but making great local LLMs is enough to get me on board!

Reply

[-]

a_beautiful_rhind@reddit

Pour one out for reasoning. Doing very little.

Reply

[-]

FullOf_Bad_Ideas@reddit

It's reasoning chain seems very short, just 300-500 tokens, similar to claude 3.7 Sonnet. It's nice to see some models with short reasoning chains, not having to wait a minute to get an answer is cool, and in my super short testing the short reasoning actually helps too.

Reply

[-]

ResearchCrafty1804@reddit (OP)

Thinking can be enabled through prompting. “Each model can answer directly (standard LLM), or self-reflect before answering (like reasoning models). This is similar to Claude 3.7, where you can pick when you want the model to answer normally and when you want it to think longer before answering.“

Reply

[-]

RMCPhoto@reddit

Any idea how this could work for structured output? Structured generation would often benefit from a reasoning step but typically requires 2 models. If this can be done with a single model that can be a game changer for local tasks. Does this support structured output natively like qwen/llama? Or has it been trained out in the reasoning steps?

Reply

[-]

Spectrum1523@reddit

What do you prompt to get it to reason?

Reply

[-]

mikael110@reddit

It's described in the readme for the models. You add "Enable deep thinking subroutine." as the first line of the system prompt. If you want to add your own system prompt as well it should be done with two newlines after that line. I've tested it out locally and it does in fact seems to work quite well. Adding the line consistently make it emit thinking tags.

Reply

[-]

Spectrum1523@reddit

Thank you!

Reply

[-]

kmouratidis@reddit

Probably also a good idea to add a thinking tag in the response before giving it to the LLM, just like the Qwen and DeepSeek suggest for their models.

Reply

[-]

dampflokfreund@reddit

This is much better than having seperate reasoning models.

Reply

[-]

silenceimpaired@reddit

It's probably in my head, but this thing sounds so much smarter than the Qwen model it was derived from.

Reply

[-]

silenceimpaired@reddit

30b is definitely not smarter than Gemini Pro for coding and made some simple mistakes, but it still sounds smarter.

Reply

[-]

dangost_@reddit

Is there are tools support?

Reply

[-]

Living-Purpose-8428@reddit

Yes

Reply

[-]

ChessGibson@reddit

Does it support tool calls ?

Reply

[-]

yehiaserag@reddit

Is this better than qwen2.5 coder instruct? How doea it do in EvalPlus?

Reply

[-]

Firstbober@reddit

Do we have any comparisons against Gemma 3? Especially multilingual tasks. As of now I don't think there is any model competing in this area with capabilities and especially the size.

Reply

[-]

Admirable-Star7088@reddit

Perfect, a 3b version too, which can be used as speculative decoding to speed up the 70b!

Reply

[-]

popegonzalo@reddit

A 10% change is not a game changer, but let's give it a try...

Reply

[-]

ed_ww@reddit

Looks super promising. I would only wish for a q4 24b param model so it fits my MacBook 32gb 😅

Reply

[-]

Federal-Effective879@reddit

I tried it out the 14B for some mechanical engineering problems and it was underwhelming, no better than the equivalent Qwen 2.5 and worse than Phi-4. I suspect some benchmaxxing that doesn’t quite reflect real world usage performance.

Reply

[-]

Leather-Cod2129@reddit

And Gemma 3?

Reply

[-]

newdoria88@reddit

How censored is it?

Reply

[-]

soumen08@reddit

I know the license is shit, but does anyone know how these compare to exaone deep from LGAI?

Reply

[-]

Key_Papaya2972@reddit

almost 90 MMLU and 75+ MMLU-Pro for a non-reasoning 32B? That's suspicious and I will test it out by myself.

Reply

[-]

Xhatz@reddit

Pretty meh in roleplay in my tests, loses coherence quickly, nothing can beat my lovely NeMo yet...

Reply

[-]

Acceptable-State-271@reddit

For reason feature Generally very good But very week for long text input

Reply

[-]

draetheus@reddit

Did some testing of the 14B model. It's not groundbreaking but I do think it is a noticeable improvement, which is more than I can say about other Qwen finetunes. The reasoning does not rely on wait so it is surprisingly short and to the point with its thinking, although I'm not sure if that makes it less robust. Interestingly, this is the only model in the <= 14B class I've seen use python type hinting without being explicitly prompted to. Long story short, the 14B model is definitely worth testing for us VRAM poor folks.

Reply

[-]

mpasila@reddit

So reasoning actually made the model worse in quite a few benchmarks except for MATH, MMLU and MMLU-Pro..

Reply

[-]

silenceimpaired@reddit

WAITAMINUTE…. I bet this is why Scout was rush released. It says on the blog they worked with The Llama team. I wondered how Meta could know another model was coming out, especially if it was a Chinese company like Qwen or Deepseek. This makes way more sense.

Reply

[-]

rb9_3b@reddit

Maybe, but due to Qwen3, not a Qwen2.5 finetune

Reply

[-]

JockY@reddit

Cogito ergo sum, eh?

Reply

[-]

sophosympatheia@reddit

The 70B Llama-based model actually ain't half bad for roleplay. It seems promising as an ingredient for merges.

Reply

[-]

GreatBigJerk@reddit

Their comparison of the 70b model against Llama 4 Scout 109b is a pretty good flex at the end.

Reply

[-]

davernow@reddit

Well it’s 70B active vs 17B active. Hard to compare.

Reply

[-]

Berberis@reddit

not sure why the downvotes, this is totally true. Scout scores 3% lower but will run 4x faster. For most, that's a totally acceptable trade-off.

Reply

[-]

GreatBigJerk@reddit

Except for the people that can actually run the slower model locally without going into debt.

Reply

[-]

Berberis@reddit

I mean, that’s fair. It’s good to have models that serve both use cases. I have a Mac Studio and would love a highly performant MoE of this size. For others, the slower dense model is better. There’s no single ‘best’ model was my point.

Reply

[-]

silenceimpaired@reddit

\* on some systems with certain quantizations.

Reply

[-]

Berberis@reddit

*all else equal (assuming you’re not vram limited)

Reply

[-]

GreatBigJerk@reddit

If I could run the 109b model on hardware that can only run around a 17b model, then I would say you had a point. The model still needs to be fully loaded to run. It's a model that came out a couple days later with fewer total parameters, and it outperforms Meta's model after they have sunk billions into it. That is a flex.

Reply

[-]

FrostyContribution35@reddit

By that logic DeepSeek is only a 37B param model. With MoEs you estimate parameters via the geometric mean. So L4 scout is approximately a 43B model and DeepSeek R1 is approximately a 157B model

Reply

[-]

davernow@reddit

So it’s comparing 70b to 43b by that measure. Either way much less compute per token. It just not an apple to apples comparison. To say cognito 70b vs scout 109b is “a flex” is a bit misleading.

Reply

[-]

silenceimpaired@reddit

I don't know... If it was llama 3.3 vs Scout then I might agree. Cogito used 3.1... 'old' llm technology, and out paced the latest technology from Meta in accuracy. The question I have is, which performs better in speed and accuracy? Q4 3.1 Cogito or Q4 Scout on my system... because if Scout fails me on both counts I have no use for it... and if that's the case... it's a hard flex...

Reply

[-]

YouDontSeemRight@reddit

In a uniform system yes, in a system split between GPU and CPU that might not hold true.

Reply

[-]

poli-cya@reddit

I mean, it is a small open source team compared to a mega-corporation with a trillion GPUs and a gaggle of engineers making millions a year to make the smartest models... Meta's 43B-alike should beat any open team's 70B.

Reply

[-]

FrostyContribution35@reddit

Regardless. Cogito 32B is still outperforming L4 scout. Also Llama did this to themselves by comparing L4 Scout to much smaller LLMs (Mistral small, Gemma 27B) instead of models in its size class like Qwen 32B and Nemo 49B

Reply

[-]

Additional_Ad_7718@reddit

Similar vam requirement regardless of active params though, all you get with the MoE is speed

Reply

[-]

Virtualcosmos@reddit

Better at math than my chinese nerd? (aka QwQ). I must personally see that to believe it!

Reply

[-]

e34234@reddit

Potentially unrelated take (as it's not about LLM tech) but I wonder why they are using Google Sans in their branding and website. It's not possible to license that typeface at all, it's reserved for Google's use. It actually made me think this was some kind of Google project. Weird.

Reply

[-]

ZippyZebras@reddit

They're related to Google like every rando in FAANG is part of the X mafia even if we leave after spending 2 years shipping a 10 pixel change. I tried the models and they're clearly overfitted to the benchmarks. Expert scam honestly: convince VCs you're going to build neo-DeepMind...even though you were on the productization side of an org *so notoriously bad at product* that they're literally paying people to sit and do nothing.

Reply

[-]

silenceimpaired@reddit

How do you know it's thinking... and how do you enable that in Oobabooga... if anyone has. I've tried to follow the model card, but I'm not very familiar with Ooobabooga's setup compared to what is shown on the model card.

Reply

[-]

Osama_Saba@reddit

Can they use tools? I don't care if they don't. I care if they don't, but don't care about the models if they don't. Do they?

Reply

[-]

xignaceh@reddit

They don't the do don't they?

Reply

[-]

Everlier@reddit

Yesn't

Reply

[-]

Osama_Saba@reddit

Oh, really?

Reply

[-]

giant3@reddit

Yeah dude, are you a native English speaker or you are translating from some other language?

Reply

[-]

Osama_Saba@reddit

Niether

Reply

[-]

shaakz@reddit

This comment gave me an aneurysm.

Reply

[-]

Osama_Saba@reddit

But what's the answer?

Reply

[-]

LostHisDog@reddit

42

Reply

[-]

Informal_Librarian@reddit

Exciting release! FYI if you want to try they are now live in LM Studio. Just search for "Cogito" and sort by "Recently Updated" and you'll see the new models listed by lmstudio-community today. In my quick tests they do indeed appear to be quite good.

Reply

[-]

noage@reddit

I'm curious why the MMLU score in the larger models is not as much (or even negatively) impacted by reasoning.

Reply

[-]

Marcuss2@reddit

Very impressive, aside from Math, it seems that the models are comparable to models twice their size. I do wonder how it will compare to Qwen3

Reply

[-]

ProbaDude@reddit

Very exciting! Will be following closely

Reply

[-]

ResearchCrafty1804@reddit (OP)

Blog post: https://www.deepcogito.com/research/cogito-v1-preview X post: https://x.com/drishanarora/status/1909672495588008312?s=46

Reply

[-]

ProbaDude@reddit

Thanks, will give it a read

Reply

[-]

FrostyContribution35@reddit

Pretty incredible performance, I'm curious to hear more about the IDA process. The blogpost mentioned techniques such as "CoT, answer verification, sampling multiple responses, etc.". Was reinforcement learning used at all in the training scheme?

Reply

Reply to Post

146 Comments