Hacker News

Feels better to me in the sense that it doesn't try to be too clever. When I ask it to do things it actually generates code that a developer would write. That still lets me build software in layers like I do mentally. And it keeps you in control of where abstractions are.

Some people would say that you don't need this at all with the frontier llms. But when extending existing software or writing software that fits into an existing ecosystem, I find I still need to offer a lot of guidance. And it's better that the context window isn't flooded with code that was reverted after I looked at it.

> My next subscription will probably be Z.ai

Wait... you just spent 3 paragraphs singing the praises of DS... and then you say you're going to subscribe to the one that makes GLM. ?

(I do tend to agree with you, though I tend to favor GLM-5.3-flash over DS 4.1 flash.)

Agreed.

It doesn’t have to be better, it just had to be nearly equivalent. I’ve been doing a lot of vulnerability work lately (trying to get ahead of the bots finding exploits in our code) and the frontier models feel like a chaperone at a church dance: get too close and they show up with the ruler and enforce the safety margin. Using deepseek and glm and qwen, I’ve gotten nearly equivalent results without the constraints. I still use frontier models for things but outside of a company environment, I’m happy to spend 1/10 the money, get the same results, and not have try and trick the AI into working.

No need to use a chinese provider, in case that's an issue for you. https://fireworks.ai provides this model for an ok price.

>> Being able to see the CoT

Dont assume that is how your model decides...

https://arxiv.org/abs/2305.04388

https://arxiv.org/abs/2307.13702

https://arxiv.org/abs/2503.08679

Absolutely!

DeepSeek feels like an excellent sparring partner. The others feel like mind readers.

I don't have to worry about any complex context engineering / multi agent stuff if I am using DeepSeek because the sequential stuff is fast enough. I don't have to bother about touching a million context length because its cheap enough.

Where it seems to lag behind is 3d, computer use, motion graphics, design but the initial demos seem like the vision model has improved substantially.

DeepSeek v4.1 Flash hallucinates like crazy.[1]

Is DeepSeek: DeepSeek v4.1 Flash 1017 any better?

[1] - https://www.youtube.com/watch?v=P4dTq4X8bqk

>My next subscription will probably be Z.ai

please don't subject yourself to this. z.ai subscriptions are not a good value. just pay API rates, or use a subscription like opencode go.

Google's Argon model via antigravity is unusable, for a small code review I asked, already spent like 10% of the weekly usage, in less than 10 mins. Very useless imho.

I always find DS or Qwen to be very spot on on audit changes or plans for just a few cents.

Is that a dig on DeepSeek not offering subscription plans?

If so OpenCode Go looks good and includes DeepSeek models (yet to try myself).

I second this, I love DS and am not looking back. For Elixir, not a lot of models are good, but DS just blazes through it.

I'm enjoying Alibaba's Token Plan. They don't have any hourly/daily cutoffs. You get your allotment for the month and use it when you want.

If I were part of a cabal that is interested in preserving control and trying to secure trillions of dollars, it would be a minor effort to nudge and bend the curve against competitors like the Asian AI complex.

I think opencode go is the better value to get access to DeepSeek.

I do not use DS for coding, but for various natural language processing tasks (even translation) in large volume, and it works wonder.

It is very capable for a lot of tasks, not only coding like the other open models.

I like it because it does the thing I want it to do rather than moralizing.

No, I don't care about license provenance for all the components of a 30 line script that i'm using in my house, for myself to sync up some folders. No, i'm not going to find and archive the licenses for all the previous projects leading to this for future legal disclosure.

Yes, I own the DVD for the movie i'm transcoding. No I don't want to talk about the legality of rebroadcast.

Deepseek (and most chinese models as a whole) don't care . oAI and Anthropic models want to make it point #1.

... which is extra hilarious knowing the laws broken to create the damn models in the first place.

there's a literal meme about this floating around the harness/AI chats for the past few months[0]

[0]: https://i.ibb.co/SwH4Sgwq/796779070-28483090404651613-920066...

> not being randomly blocked for cyber review

Just to test, will GLM – hosted by an infra provider other than Z.AI – also stop answering if I asked controversial questions from the pov of its country of origin? Or third party infra providers can perform some unhinging on the model before hosting? Or actually the blocking happens even before the question reaches the LLMs in that country?

(Note: my paid LLM access at this point are only via DS and GLM subs - 90-95% usage; and a free Gemini Pro access. I hope that adds some cred that I am not trying to take potshots at the their country of origin).

Deepseek and Muse are my go to for cheap. both work perfect for basic and mid tier tasks

> Deepseek... My next subscription will probably be Z.ai

Wait wut? You do realize that they are two different companies, right?

and really , for 90% of standard agentic work, deepseek flash is more than good enough. its the main driver for our Personal Assistant offering and users love it.

> Being able to see the CoT and not being randomly blocked for cyber review.

I had a bad experience with MiMo from Xiaomi. I had it trigger unsafe content blocks while fixing visual bugs in a completely vanilla space game it just wrote from scratch in the same session. It also got triggered for changes in some other benign greenfield app.

For us neophytes, what does this mean: “randomly blocked for cyber review”?

Agreed, but don’t subscribe to z.ai; their inference is terrible. Stick with openrouter

Isnt Z.ai for GLM? Does Z.ai provide Deepseek?

Huh, doesn't DeepSeek take the longest because it overthinks way too much? Haven't tested this release yet but was very disappointed with past ones because it takes an eternity to do anything because it constantly overthinks and questions itself way too much instead of just starting.

I love using glm 5.3 more than deepseek because of performance

also it's cheap, using it in opencode is great

For 99% of the tasks any model really would do, even GPT 5.4 from an year ago.

It's only for long running/very difficult cases that the differences are here, and they are small.

still a myracle to me what people achieve with it? I tried so many things with it, everything is broken it produces

Their strategy is regulatory capture.

DeepSeek is simply amazing.

100%

I keep seeing this assertion. What regulations have they proposed that would have prevented this new model? I thought they want to pace the frontier, which this model is not.

We'll put that link in the toptext too. Thanks!

LLMs are non-deterministic, you could run the same prompt on the same code base with the same models, and have different result. I don't think only one run with each model is enough to draw conclusions.

you probably tried with DS on max and I don't blame you.

But try it on high instead of max if possible. max is known for being overly verbose to the point of being detrimental compared to high at times.

For example Simmons svg pelican took 11 minutes on max and 1 minute on high. And they are comparable.

In your case, is it a public repository that the llms might have found the issues for online?

I too am thinking about creating some benchmarks that I can understand myself. The LLM 'cheating' by finding known issues through some sort of history or online source is a path you want to protect against. Then again, you DO want them to be able to look online for people having similar issues.

> Worth to mention: none of those findings overlapped.

This is why using multiple model families is important. Don't limit yourself by being a stan for this or that company.

> Worth to mention: none of those findings overlapped.

Yet to be obsolete.

What is p1, p2?

This read like a conversation you were having in your head -- unfiltered.

Was that the old DS 4.1 Flash or 4.1 Flash 1017?

Flagship gpt models are crazy efficient. 6.1 Sol is like a savant that looks at a problem and spits out the answer seemingly without thinking.

Opposite of my experience - openai models, and codex are clearly the slowest (out of 5 different models/harnesses I use on a regular basis).

Kinda shows their team is more engineers than marketers.

To me this implies they continued the training run for 4.1 and it improved. A 4.2 would imply they changed something about how it’s trained/how inference works, and a 5.0 would be a major update (who knows what would have changed).

This is just my headcanon though.

It's not really contradictory, minor changes can have major effects.

That's one small step for a man, one giant leap for mankind.

Hoping it's minor relative to what's coming

Minor updates can provide substantial improvements sometimes, it really depends on the specifics.

They changed how it works substantially. I think it should be 4.2.

Quiet confidence backed by provenance :).

It's a bit refreshing to read updates from an AI lab without bullshit doomerism and overwhelming sense of self-importance from shit like OpenAI and Claude.

Also, DS is an excellent model.

I'm at 1 trying to get to 10. I move to 2. Minor update. 100% improvement towards goal.

These are not mutually exclusive.

That is how exponential functions work. Improve every aspect of a system with 1% and you may get easily 50% more performance.

It's about playing to two different audiences.

The CCP requires extensive certification for new major releases, to make sure a model doesn't do anything they don't want it to.

To get around this, Chinese labs only bump versions when they have to and usually claim it's a minor update.

Then on the other side, they're obviously trying to market to the tech crowd...

Improved on exactly the items that I hear people complain about with the previous checkpoint!

Ok but this is a new model with significant architectural differences and new training run.

I'm baffled, my $20 is enough for me to do some consulting, personal projects and side work, and still cannot outran my quotas, I'm building different kind of things: flutter apps, dotnet backend, database, front end, architecture diagrams, code exploration, tests, PowerShell scripts.

How people spend more than 200 usd monthly is beyond my understanding, are you guys so inefficient with the usage? or I am the opposite?

> I run through a $200 Claude Code subscription in less than 24 hours now just for a single project

How?

> I run through a $200 Claude Code subscription in less than 24 hours now just for a single project

I would be worried about your methodology when the subscription party ends. Make no mistake, one day, the party is ending and the people who are dining out on approaches that burn massive tokens are in for a rude shock.

After opus 5.5 release I find it very hard to run out on a 100 sub...

I'll bite. Wtf are you working on that takes $200 a day and you're willing to pay for?

Interesting, coming from a similar situation, I'm interested what workflow you landed on. Opus planning + available to be called for deep thinking + opus reviwer? or?

Yeah the Claude/anthropic usage is quite arbitrary, I also run through my entire 20x pro plan in 1-2 days. Codex has slightly better usage levels, but is annoying to use for many reasons. DeepSeek flash (direct or via opencode go) is the best value model I've found so far.

This is a new model.

I have quite the opposite experience: I can do any concrete coding change with any model, but the intelligence limit just doesn't go through more complex debugging examples.

And then when I try a smart model, like Astra medium and it solves it (I had a firmware bug in my new laptop where I couldn't charge through USB-C) on the first try (altough with a repo already having lots of information about the hardware/firmware/debugging programs/traces), I realize how much time I'm losing on the few examples that the less intelligent models are just not able to debug and solve.

Yeah, but for some reason people think they need SOTA to help them work on their side/indie projects.

Like if we needed PhD intelligence for productivity tools and SaaS platforms, why weren't we hiring PhDs to build things before.

99.99% of people do not need SOTA ai.

Speed and price. You should try the 4.1 Flash on one of the fast providers. It's pretty crazy how fast this model works, and it's great for iterative software development because you can steer it fast to where you want to go.

https://www.lithosai.com/

https://www.coralbricks.ai/

Same. I'm just happy sticking with them and am consistently impressed. Cost is so good and kind of keeps you honest.

definitely could use less thinking time and lower token usage, which is what this bump is focused on

> improve on what's already there for any task I know of

What about "good governance"?

Does it have zero hallucinations? Does it reflect thoroughly on every single idea it uses, verifying its solidity?

People will pay for more intelligent models. Astra and Fable are more intelligent than these open weight models and can solve problems that are much more difficult, and that's extremely valuable. It's not any more complicated than that.

I don’t know what I’m talking about but it seems to me OpenAI is heavily emphasizing productization of LLMs. Dots, ChatGPT Work, scheduled tasks, in-app projects etc.

Being able to directly distill their huge frontier models into Luna/Haiku-level models puts them a step ahead still.

i hate dario's politics, but i still find claude's eco system easier to use. Design/Artifacts is very convenient for people like me who has zero UI/UX expertise, and claude code's remote session works seamlessly on ios with a well-design client interface while you have to deal with ssh and typing in cli with your iphone keyboard if you are using opencode/omp etc

And yet, nearly all of enterprise money spent on AI tokens in the US is going to either OpenAI or Anthropic.

That's why they are constantly fearmongering and trying to get other models outlawed. If nothing else, under the pretense of national security.

v4 was a worse overthinker. 4.1 mostly fixed that. It still prefers to plan everything before making a single file change.

If I'm impatient I switch to Luna for a single turn just to get it started and let DS continue.

What is your workload?

Try the Qwen 3.8 flash next model on DwarfStar - you will be very surprised. I’ve been using the 4 bit quant almost exclusively after having DS v4 flash be my daily driver. Qwen is faster and seems to be much better at coding tasks.

This is a new model, not the same as a different one.

For comparison, close to Opus 5. Far behind Sol 6.1 and Opus 5.5.

Probably a pretty decent model for 90% of the usecases!

I've been doing a lot of hobby work with GLM 5.3 Flash and I can count the amount of times that I need a stronger model on two hands.

I have access to unlimited tokens at works and Opus is great but also often overkill if you know exactly what you need.

By next year we will likely see GLM 5.3 level performance in Qwen 4 MoE models that can run on most higher end consumer systems.

I don't think there is a reason to care about any of it, just get a few dollars on a few openrouter or similar platforms and just use the models for now, wait for cheaper models that are good enough for you to become available for your hardware.

Sure you might never see frontier performance on consumer hardware but do you really need that though? Does anyone? And if you ever do just get a 30-day sub for that work, and cancel it.

Cant you run it on a pair of DGX spark? Sure it's expensive now, but at the initial retail price that was what, 8k in hardware? Not exactly unachievable

Forgive the ignorance, but how are you running a 763B parameter model on 64GB of RAM?

You should start asking for animated three.js pelicans too, the 2D pelicans seem to have saturated.

This is in "high" mode .. 40 seconds only: https://ibb.co/5XTRKNJv

Looks much better.

Wow. Is DS the only model that consistently puts one leg behind the frame? I don’t even remember Opus ever doing that.

The best bicycle is "low".

You trust AA since the sol fiasco a month or two ago?

Why not just try it yourself?

I may be missing something so apologies if I'm telling you something you already know but: there are plenty of web search APIs for agents - the Brave one is the one I use: https://brave.com/search/api/

I gave it a kick and got mixed results. For example, searching for a Microsoft product couldn't find its landing page, and searching for a niche open source package landed in the Debian repos for the top hits, but the home page was in the first 10 hits.

Lovely to see they're doing their own search though, I love how many folk are actively working on search clones at the moment.

Already released https://api-docs.deepseek.com/guides/search/

There are theoretical attacks involving malicious weights. But the risk profile is so outrageously in favor of them not doing it that you'd essentially have to be an anti-China conspiracy theorist to find it plausible, let alone likely.

Registered just to ask you.... 50 instances? Which harness? How? Why? Can you post a video demonstration?

Could you please share your methodology?

Exactly, there is no ”best.”

I've been telling people this for a couple of months now, I'm actually struggling to use up all my tokens.

But this is a new, different model that just came out.

What does it matter if you’re not hitting your cap with OpenAI/Anthropic in the first place? This thread is apparently populated with the highest of powerusers running multiple OpenClaws. Us peasants aren’t concerned with meters or cost/token - just quality - and nobody comes close to OpenAI or Anthropic in that regard.

Well, at least mostly everyone outside the US.

Anyone that has observed China since the 90s.

There are different types of freedom and China's bet is that as long as they provide enough economic freedom the public will be ok with somewhat restricted social and political one.

That's an image China wants the rest of the world to believe. Towards inside it still fits the typical stereotype.

that's honestly insane. it's like the go-to for when you don't want to use opus 5.5/sol 6.1

There's basic reference inference code released with the weights, yes.

https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/...

> providers have to implement which could be different than what the labs themselves use.

I'd imagine (most) providers just take implementation by vLLM/SGLang.

Yeah, I have the same question. I'd like to incorporate cheap & fast models like DS into my workflow to reduce my reliance on OpenAI/Anthropic, but after trying several such models and having to have a frontier model review/fix/redo everything multiple times per task, I gave up and simply have the frontier models (Sol/Opus) do everything. All of this fits pretty comfortably in my monthly subscription.

Maybe the equation changes if I exceed my subscription usage, and need to fill the gap with something else. But it seems like a huge waste of your time to go out of your way to use a cheap model if you have the tokens to spare in your subscription.

You can stick with Claude Code if you're already using it. DeepSeek has an Anthropic-compatible endpoint, so you just need to set ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic and ANTHROPIC_AUTH_TOKEN to your DeepSeek API key.

Pi is very good with open weight models (proprietary ones too)

Pi if you're a minimalist

OpenCode if you want a good, opinionated setup

Yes. It’s actually free to use at the moment. Of course with a quota.

Presumably. The current opencode slug for the model is just 'deepseek-flash' suggesting they will just update that one going forward, not keeping versions in parallel.

and do we know which version runs on opencode at the moment (on the free and on the paid version), is there a way to find out?

Isn't there a lot of overlap between PLAN, PHASES and DECISION that pollutes the context ? Whereas a simple PLAN.MD to put the harness on track and an ARCHITECTURE as a single source of truth to prevent it go offtrack would suffice and gives it more space for higher level of reasoning ?

Have you thoroughly tested a slightly simpler workflow with one less stage but a frontier model doing the initial planning? Luna is amazing but it's a deliberate trade off designed for speed and cost savings. You can also give them a rule about speaking clearly.

Pretty cool, thanks for sharing.

What crash?

I really like pi, though it does require some time in setting it up. But it has greatly reduced my token spend. The relevant Deepseek Harness https://www.deepseek.com/en/harness/ I hear is also good and popular.

I use deepseek via fireworks using Claude code.

When I type deepseek on my CLI it launches Claude with my fireworks API key. Works very well and feels almost free.

PSA OpenAI&Anthropic train on your session if you use their endpoint and/or harness. I'm an American! so I'm fine, but you may not be, so be aware

Presumably they registered, but that doesn't mean the model passed through export review. It just means that some bureaucrat now has the opportunity to check the product characteristics they reported and decide whether that kind of thing should be allowed for export or not. (Though with the weights out in the open, that's maybe a bit pointless.)

At this rate the frontier labs will be Anthropic and DeepSeek, with OpenAI struggling to keep up.