Talking about open models at #TRG26 🔥 @barckcode as the real boss 😎

Open models have gotten really, really good. So why are so many companies still sticking with closed ones? DeepSeek, Qwen, GLM and Kimi can already compete with (and often beat) the last closed models most companies use in production, like Sonnet, Haiku, Terra, Luna or Gemini. And they can do it at a fraction of the cost. Yet plenty of teams are still hesitant. We've been thinking about why, so we wrote a piece on it. A few things worth looking at: → Price per token doesn't tell you much. What matters is how much you pay for a task that actually gets done right. DeepSeek V4.1 Flash's latest update brought that cost down 54%, without changing the price. → DeepSeek isn't just opening its models. It's also releasing the inference engine and harness. That's a pretty big deal when you consider how much of this used to be locked inside the labs. → "We don't want our data going to China." Fair concern. But using a model trained in China doesn't mean your data goes there. It depends on where you run it and who's operating the infrastructure. → "We don't want to manage GPUs." You don't need to. Just like you don't need to run your own database servers to use PostgreSQL. Or your cloud. We're still treating open models as if quality were the main obstacle. For a lot of real-world use cases, it isn't anymore. The questions companies need to answer are where the models run, who operates them, and how much control you have over the whole thing. And those are choices companies can already make today. Because with open models you have control. Full piece: helmcode.com/posts/open-vs-c…

🐳 @deepseek_ai V4.1 Flash 1017 is now on Helmcode too. Same size, same architecture. DeepSeek focused on getting more out of what the model already had, and the gains are impressive. From DeepSeek's model card: → Terminal-Bench 4.0: 31.2 → 42.3 → ProgramBench: 20.3 → 29.1 → Automation-Bench: 54.8 → 60.5 → DeepSWE v1.1: 74.2 → 77.0 Artificial Analysis points the same way. Its Intelligence Index goes from 39 to 45, and the average cost per evaluated task drops from $0.49 to $0.26 on its test suite. Around 47% less. We also wanted to see what it does with a real project, so we ran two tests on design and video! 1/ Signal / Noise. Same creative brief for DeepSeek V4.1 Flash 1017 and Claude Opus 5.5: design and build a website from scratch. Both got to a good result. Pretty impressive advance from DeepSeek, to be honest. You have a video comparing both solutions. 2/ A motion graphics piece about what happens when machines begin to dream. Built entirely in code, with no video generation model. Opus 5.5 level on video too for a piece of cost. Improvement from 1017 is savage! Haven't tried it? Give it a full project, let it use tools and check what it built, how long it took and what it consumed. You can use it with flat rate, zero-prompt retention and EU servers at Helmcode :) Full write-up and the video: helmcode.com/posts/deepseek-… Both websites, screen by screen: helmcode.com/design-challeng…

He oído banco de tokens?

🐳 NEW DeepSeek V4.1 Flash 1017 is now live on nan.builders! Impresive post-training/update from one of oure most loved models. Improved for coding, agentic workflows and design with better reasoning. - 1M-token context window - Multimodal: text and image - Thinking mode Soon on Helmcode Enterprise plans too!

Tomorrow at AWS Community Day Spain, @crstiax explains how to size an LLM deployment on your own infra. Weights are the easy part. KV cache and batching decide whether a GPU serves 2 users or 20. And the secret sauce about what we are doing at Helmcode. Sat 17 Oct, 16:10, Santiago de Compostela @awscommspain @AWSCommunity

Which open model should you use instead of Claude, GPT or Gemini? Clients ask us this every week, so we built an equivalence table: which open model to test first if you’re using Opus, Haiku or Sol today. Benchmark rankings alone aren’t a source of truth, so we compared closed and open models by class and by workload. What you’ll find in the article: helmcode.com/posts/open-weig… An equivalence table: closed → open Benchmark numbers and where the models are close The best open model for each workload Routing as an architecture to save time, money and resources In short, after these last few weeks, this is our starting point. Not gospel, just guidance: Reasoning and knowledge work → MiMo V2.6 Pro Frontier coding and agents → GLM-5.3 Agentic coding with a better balance → GLM-5.3 Flash Tools and automation → DeepSeek V4.1 Flash High volume, basic RAG → Qwen3.6 / Qwen3.8-Flash-Next Long context → Kimi K3 Opus and Astra still win on certain problems, but we’re keeping an eye on the next open releases 👀 The setup that works best for us is routing each request to the smallest model that solves it well, and escalating to a frontier model only when it’s needed. In six months some of these names will have changed. The method won’t: take a few hundred real tasks, run them blind against your current model and two or three candidates, and measure cost per completed task. Benchmark your own workload before deciding anything.

Last week we ran our first conference. Here’s a short recap of Kernel Panic, one week later. We filled the Auditorio Casa del Lector at Matadero, Madrid. Every seat taken, people standing at the back, and nobody left early. Four talks on running open AI at scale in production, with links to their slides: → David Yunta @YuntaDavid and Luis Valerio (Inditex Tech): Kumoss, the Inditex Tech platform for building infrastructure with AI, made for developers inditextech.github.io/kumoss… → Sergio Paniego @SergioPaniego (@huggingface): Training a coding agent through a harness you did not write huggingface.co/spaces/FineEn… → Sergio Guijarro (@Cloudflare ): How Cloudflare rolled out agents to everyone in the company kernelpanic.sguijarrodepoo.c… → Cristian Córdova @barckcode (@helmcode ): How to scale inference for 1,000 users helmcode.com/slides/kernel-p… To everyone who came, thank you. You showed up, you asked hard questions and you stayed for the conversations afterwards. That’s the whole point of these events. To Hugging Face, Inditex Tech and Cloudflare: thank you for building this with us. Pictures at kernelpanic.helmcode.com See you at the next one. Because there will be a next one, and it’ll be bigger :)

🕹️ Your AI agents shouldn’t stop working because you closed your laptop. NaN Workspaces are live for your personal subscription. We’ve been building the best open-model AI subscription out there at nan.builders: good models, billions of tokens, a few of them unlimited. The missing piece was somewhere to actually run your agents. Until now, that meant renting a VPS, setting it up yourself and babysitting it. A Workspace is your own Linux machine in the cloud, with Hermes, Pi and gentle-shell preinstalled. Open it, start an agent, close the lid. It keeps running. And it’s the same session everywhere. Kick off a long task from your desk, check on it from your phone on the train, pick it back up at home. Nothing to reconnect, nothing to redeploy. Leave a scraping agent running overnight, read the results over coffee. Let a coding agent grind through a refactor while you’re in a meeting. Run a research pipeline all weekend without your laptop open. Your first Workspace is included with NaN Premium. On the regular NaN plan, add one from €4.99/month. nan.builders/docs/workspaces

You can use DeepSeek without sending a single token to China. Same for GLM and Qwen. We download the weights, deploy them on our own infrastructure, and serve them without a single request going through DeepSeek, Z.ai or Alibaba’s APIs. At Helmcode, we run them on European infrastructure: → prompts and responses are processed in the EU → the model developer never receives your data → we don’t store prompts or responses → no international data transfer for inference → you can sign a DPA with a European provider We've created a page explaining every detail helmcode.com/security/open-m… And there’s another big part: cost. - Running 10B tokens/month through Claude Opus 5.5 via API is around $72,000 at list price. - Our Growth plan is €1,299/month and includes up to 15B tokens across our open models. It means running a strong model in production no longer has to mean sending your data to a US or Chinese provider and paying frontier-model prices for every token. Your data is yours.

We are partnering with @lobera_ai and some other companies to bring AI infrastructure for the next generation of hackers and security engineers. We are proud sponsors of cyberchallenge.ai/en/ Cyberchallenge is a Capture the Flag league of Spanish universities. Students won't just hack. They'll build agents that hack. These people are building the future of security, and we want to be part of it. We're giving every team the AI infrastructure to do it: open models through the Helmcode API for teams. We'll be sharing the league board as the competition unfolds. Follow along: cyberchallenge.ai/en/

Kernel Panic! is one week away. Really short: Oct 6 · Matadero, Madrid · 16:30 4 talks on AI at scale with @huggingface, Inditex Tech and @Cloudflare. Sold out. But we just released 40 last-minute seats. First come, first served: kernelpanic.helmcode.com

🚀 Qwen Image 2.1 is available now on Helmcode. And it's a beast! 👇 → Generation and editing in one model. Same endpoint, no checkpoint swap. → Native 2K output (2048×2048), no upscaling. → Text that actually reads inside the image. Posters, packaging, UI mockups, slides. → Transparent PNGs straight from the prompt. Real alpha channel, no background remover. → Up to 10 reference images for edits, with masks or painted annotations to say exactly where. Runs on EU infrastructure, zero logs, OpenAI-compatible API, as always. Point your base_url at Helmcode and you're generating. helmcode.com/models

📸 Missed the first Café Helmcode? You're going to wish you hadn't. Here's a recap video of how it went. Lots of people sharing, lots of people having a great time, and a community we're building around open models that left wanting more. Spot any familiar faces? Thanks to @adrilaborda @iamzenitram and @cangri2k5 for sharing their projects with everyone and sparking the conversations. And a huge thank you to @Saulgmz_ for shooting this amazing video. The GOAT 📷 There'll be more. Many more! Coming to the next one?

Yesterday we hosted our first Café Helmcode Madrid 🤩 A meetup about AI and open models for +120 people at Mad Tech Campus. Just 3 lightning talks plus good networking for the best community :) Was a nice and warm event. Community. AI. Good talks. Good people. There will be more. Next city?

Today is a special day: we're introducing the first 4 NaN.builders /Helmcode ambassadors! Who are they? Four absolute legends: - @686f6c61 - @Saulgmz_ - @juan_barbat - @cangri2k5 Serious token burners who also bring knowledge, help, resources, projects and a lot of teaching to the community. NaN community is growing fast, and these four have been with us since day one, build really cool stuff and spread the word. Now they're going to give us a hand. Each from a different angle, all with the same goal: bringing open AI to every corner of the planet. They'll help us set up Cafés Helmcode in other cities, run workshops, share how to get the most out of the NaN subscription and answer questions. The NaN Ambassadors program starts now. Over the coming months we'll open a public call to pick the next 4 by the end of the year. Give them a warm welcome and follow them, you'll be hearing a lot from them :)

Tomorrow, 22nd will be hosting our first Café Helmcode in Madrid. An event about open models and people building and learning with AI. We'll have 3 short 20-minute talks to get the conversation going: - @adrilaborda , SWD at Spain's Ministry of Interior, on "AI in public administration" - @iamzenitram , infra at RevenueCat, on "Monitoring football blackouts" - @cangri2k5, Santander Bank, talking about data grounding in model fine-tuning The event has been full for weeks, but we'll be hosting more Café Helmcode in other cities soon :) And we will show you some pictures of the event too

Sometimes, the best proof of trust you can show is just an email from a big client with this kind of compliment 😎 Best DeepSeek inference performance? They say it If you are looking for an open model provider with dedicated infra and zero logs, we are your company helmcode.com/contact

What happens when you write a prompt? Where all the context goes? What your inference/model provider does with all your data? Nice questions to answer in our latest article: - We talk about what travels with your prompt - Understand the context first - What wording matters to you (not for training, zero-data retention, dedicated inference) - And how to establish boundaries for your workload Read it here: helmcode.com/posts/where-you…

Tickets are now open for Kernel Panic! Helmcode's first open AI conference 😍 📆 October 6 📍 Casa del Lector Auditorium, Matadero | Madrid 🕛 4:30 PM (doors and check-in) to 9:30 PM 4 talks with open Q&A on real AI projects at scale 👇 - From specs to production: Rethinking cloud infrastructure with AI | Luis Valerio and David Yunta, Inditex - Serving models at scale for 500+ people | Cristian Córdova @barckcode, @helmcode - Training a coding agent through a harness you didn't write | @SergioPaniego, @huggingface - TBD, Alba Massa | @Cloudflare - Closing and afterwork Capacity: 200 people. Approval required to attend. Grab yours here: luma.com/p50cydsf

All you need to know about DeepSeek-V4.1-Flash - We did a deep dive into its new architecture. Worth reading: helmcode.com/deepseek-v41-fl… - If you want the shorter version, our article on the launch: helmcode.com/posts/deepseek-… - It's now available on Helmcode, with 5B/15B/35B tokens depending on your plan

Inside DeepSeek-V4.1-Flash 🐳 Yesterday @deepseek_ai dropped a new Flash model with a new architecture. But what actually changed? What is that architecture? And why does the KV cache matter so much? So we're publishing our technical analysis of the model. We spent some time reading their paper and running experiments on DS-V4.1-Flash, and put together a deep dive on everything you need to know: - The new architecture explained: CED, CSA2, FP4 - How the cache is handled now - Costs, results, and what's different from V4 Deepdive here: helmcode.com/deepseek-v41-fl… Easy to read, easy to navigate, and easy to understand. We built custom figures to illustrate the harder concepts. Hope you like it!

Khé es esto? :O watafack

"Unlimited tokens" sounds like a trap, so here's the actual math. 500 users don't consume 500x the capacity of one, because they don't spike at the same time. Inference at scale is statistical multiplexing: the same physics that makes your ISP work. Batching fills the gaps; the GPU rarely sits idle and rarely saturates. That’s why we can offer a hight cap of tokens on SoTA models too. Our margin doesn't come from your restraint. It comes from our optimization: that's why we publish the engineering. For example: +11.3% throughput on the same GPU with one quantization change. helmcode.com/posts/qwen36-rt… The meter for token caps is a billing choice.

Legal work has a property that makes most AI setups a non-starter: the documents ARE the confidential asset. There's no anonymized version of a contract under negotiation. What legal teams run on open models: clause analysis with 1M-token context (the whole contract, not chunks), document triage by matter and jurisdiction, first drafts from approved clause libraries. And the meta-angle: legal teams are also the ones asking every vendor where inference happens, because the AI Act made that their job. Our answer fits in one line: EU infrastructure, zero logs, auditable stack. How legal teams deploy this: helmcode.com/industries/lega…

We rebuilt Granola with open models. 👀 But the meeting notes turned out to be the least interesting part. Every customer call contains decisions, objections, feature requests, promises, problems... Most of that intelligence disappears the moment the call ends. So we built a pipeline to turn calls into searchable memory: - Whisper large-v3 → transcription - pyannote → speaker separation - DeepSeek V4 Flash → summaries, decisions & action items - Qwen embeddings + reranker → memory across every meeting The result isn't just summarize a call. It's memory and intelligence: - "What objections keep appearing in sales calls?" - "Which features have been requested by 3+ customers?" - "Why did we kill that idea six months ago?" Your calls are already a company database. You just don't have a query layer. This is the first AI Use Cases 101: practical recipes to build real tools with open models. Article with the details, here: helmcode.com/posts/build-you… We published the full architecture, models, lessons learned and the repo. You can also fork helmcode-whisper and build your own instead of starting from zero: github.com/helmcode/helmcode… Own the audio. Own the memory. Own the intelligence.

This is the resource you need for you daily work. What is the best open model for every use case? Our team mapped 80 use cases paired with the best open model for every case. We took 80 real enterprise use cases (from summarizing email to due diligence in a data room) and matched each one with the open-weight model that actually solves it: the recommendation, the alternative, and the why. Find your model in two questions: helmcode.com/open-model-guid… How it works, in 3 points: 1/ Quick selector: pick your task type and your main constraint (quality, cost, single GPU, long context). One recommendation, instantly. 2/ Every card shows its reasoning. Not "best model" but why THIS one: "1M context swallows a full report in one pass, no chunking. Bigger windows exist in worse models; the window only tells you what fits, not what the model understood." 3/ It's a guide, not a catalog. It recommends plenty of models we don't serve: Voxtral for transcription, Gemma 24B for edge classification, Kimi K3 for data room reasoning (which needs a 64-GPU cluster and we tell you that too). Visit the guide and save this resource or share it with your team! helmcode.com/open-model-guid…

This is just one example of what people in our community are doing with open models. We don't only serve companies. There's a community of more than 500 builders using open models for their own projects. @686f6c61 built a radar on @huggingface to track every new model on the platform, using his nan.builders subscription. He's also a NaN Ambassador. Check out his project here: radar-huggingface.686f6c61.d… Model trees ara just beutiful 😍

Por alusiones ;) Si es infra de IA, gestionamos y operamos infra de otros Y montamos, gestionamos y operamos on-prem también

GLM 5.3 is now available on Helmcode plans!! 🚀 We were testing GLM 5.3 and it impress us. Is the biggest and better open model efficient/cost right now. And is live on Helmcode Enterprise accounts. The most important thing is this: - Benchmarks as Kimi 3 - At a fraction of his cost! Good post-training on coding and long-horizon tasks by @Zai_org We serve it form Europe, on EU infrastructure, with zero logs. Flat pricing, built for real workloads: pick your tier (starter, growth or scale) and add the API keys you need with GLM 5.3. €150/month per key. Volume comes from your tier, as a monthly token allowance shared across the account: 2B on starter, 6.5B on growth and 18B on scale. No overage charges, ever. If you are wasting more than 100M tokens on GLM 5.3 right now, our flat fee is a better price. You can see it in third image. We explain more about the model and capacities and compare it with Kimi and a pricing table on our blog: helmcode.com/posts/glm-5-2-n… And a little gift for all of you. All the Enterprise accounts had Deepseek-v4-Flash with the last version with vision included. Go to your dashboard and add your keys or come to the dark side and be part of Helmcode and start burning tokens! 🔥

Today we launch our Open Model Guide for Enterprise Use Cases 👇 Too many models. GLM. DeepSeek. Qwen. Kimi. Gemma. Nemotron. MiniMax. Maybe two more will be shipped this week. Access to open models stopped being the problem a while ago. Choosing the correct one is the problem now. Which one do I use for my task? Do I use the same one for everything? Which one is best for the other things I do? Nobody has time to benchmark eleven models against their actual work, so most teams pick whatever topped a leaderboard once and route everything through it. And that’s a bad idea… So we did the work: we've created a guide with 80 enterprise use cases, each mapped to the open model that actually solves it, with the reasoning and the alternative on every card. 👉 Open models guide for every use case: helmcode.com/open-model-guid… You can download it in PDF too. How it could work for you: 1/ There’s a selector with two questions, and our recommended model. Tell it what you're building and what matters most (quality, cost, single GPU, long context) and you get your recommendation. Ten seconds. 2/ No, you shouldn't use the same model for everything. Seven open models cover 95% of use cases. You don't need a model zoo: just a short list with clear jobs. 3/ The leaderboard winner is rarely the right answer. Benchmarks don't measure real work. Two index points cost 3.5x more per task. Five points cost 31x. Your ticket classifier will never notice the difference; use closed models for your quarterly market analysis. 4/ Every card explains WHY, not just what. What decides a recommendation is rarely the score: it's visual tokens per image, real context vs advertised context, GPU nodes to serve it and license. 5/ The green dot marks what runs on our platform. Everything else is still the right call. Free, EN/ES, PDF included, no signup. Look for your use cases and start using the right model!

🗓️ September starts tomorrow, and with it the season of questioning subscriptions. This year there's a new line under review: the closed-model AI contract signed back when there was no alternative. We're getting more enterprise questions than usual these weeks, so let's answer the five we hear most: 1/ "Can open models actually replace what we run on closed ones?" For most workloads, yes, and we don't say it on faith: 99.5% of the tokens we serve in production run through open models. Classification, extraction, RAG, coding, agents. The gap with closed models only exists at the very edge of the frontier and almost no real workload lives there. We mapped 80 use cases to the right open model, free: helmcode.com/open-model-guid… 2/ "What does flat-rate include? Is unlimited real?" You pay per API key, not per token. Unlimited tokens on open models, plus per-model allowances for SOTA-class ones (not shared between models). There is no overage mechanism, by design: it's technically impossible to spend more than you contracted and rate limits on SOTA are big enough. 3/ "How painful is the migration?" It's an afternoon. The API is OpenAI-compatible: same SDK, one base_url change. Pick your most expensive workload, run it in parallel, compare outputs. And then, migrate everyhting. 4/ "Where does our data actually go?" EU infrastructure, and we don't store prompts or outputs. Your auditor gets a one-line answer and the documentation to back it, which matters more every month now that the AI Act's due diligence questions are landing in procurement. 5/ "And when we still need GPT, Claude or Gemini?" Keep them, through the same API: frontier models run on prepaid credit, your plan is consumed first one integration and one invoice. A recent client kept closed models exactly where they add value (R&D) and still cut their AI bill by 75%. The honest caveat, as always: if your volume is low and your data can travel, a metered API is cheaper and you should use it. Flat-rate sovereignty is for teams with real volume, real data constraints or both. 🌞 September is a good month to run the numbers: helmcode.com/pricing

Mañana no, la semana que viene que aún tenemos que optimizar bastante xD Pero para que te hagas una idea, comparando los €70 de NaN con precios API de hoy te salen €1400 :P Solo con GLM, vía API te gastas más. Pero deja que carguemos bien el modelo :)

🔥 Yesterday two brand-new open models shipped both were LIVE for our community the same day: - Qwen3.8-Flash-Next (1B tokens/mo) - GLM-5.3-Flash (500M tokens/mo). Day-0 support, already usable. If you haven't heard of NaN: it's our builders community, and it's also the answer to "does Helmcode have a personal plan?". It does, and it comes with 500+ builders attached. What €70/mo gets you: → Qwen3.8-Flash-Next: 1B tokens/mo → GLM-5.3-Flash: 500M tokens/mo → MiMo v2.5: 1B tokens/mo → DeepSeek V4-Flash 0731: 2B tokens/mo → Qwen 3.6 + Gemma 4 unlimited (plus embeddings, rerank, whisper and somo more) → The Discord: workshops, hackathons, quarterly votes on which models we serve next and IRL events like Café Helmcode (fewer than 15 tickets left for the first one) → All of it with zero logs stored, on EU infrastructure We just opened 50 new spots on the waitlist. Join today, approved today: nan.builders Come burn some tokens :)

The paper behind the new DeepSeek Harness 👀👇 Nobody reads an 88-page programming languages paper. We did, so here it is in plain language. "A Programming Paradigm for Spatiotemporal Composability" is the paper behind Cordis, the plugin framework DeepSeek Harness is built on. It answers one question: how do you compose software that changes while it runs. Read the whole piece here: helmcode.com/posts/cordis-pa… Two ideas carry the whole thing. -Revertible effects, the time half. Everything a plugin does on load carries its own inverse, so unloading it leaves nothing behind. No orphaned listener, no open connection, no ghost command. -Reactive coeffects, the space half. What a plugin needs is declared up front, so the runtime will not start it until those needs are met, and reacts if they disappear later. On top of that, five proven theorems. The one that matters is confluence: hours of loading and unloading converge to the same state as a clean install. The order of changes does not change the destination. Composability stops being an architectural opinion and becomes something you prove. Worth knowing even if you never touch DeepSeek Harness, because this plugin pattern is spreading. OpenCode v2 is in beta with a rewritten plugin API where setup returns its own teardown function. Same idea as revertible effects, arrived at independently. We made the work: translate the paper to plain text so everybody (in tech) can understand. Full read, free, no signup: helmcode.com/posts/cordis-pa…

Solo podemos decir esto :)

Claude down again? Wooof 😓 We are seeing a lot of Claude users cancelling their subscription. And you are right. Come to the dark side. Stop paying per token and stop being interrupted while working. We’ve raised our Deepseek v4 Flash 0731 limits to help companies and builders work with open models. - Companies: €1,299/ month for 15B tokens of Ds4 (and unlimited Qwen 3.6 and Gemma4 and more). Up to 15 API keys 👉Visit helmcode.com/pricing - Builders: €70/ month for 2B tokes of Ds4 (and unlimited Qwen 3.6 and Gemma4 and more) 👉 Visit nan.builders Stop paying per token. Feel the freedom of the flat fee.

Our @deepseek_ai Harness Guide is getting notice!! This week, more than 5K people 🤩 around the world have visited the microsite to learn more about the harness and its architecture. What will you find inside? - A comprehensive guide with 23 sections explaining the harness and the paper published - Cordis and plugin architecture explained - 4 reading paths: from beginner to pro with setup to theory - Deepseek whale diving on your screen :) Learn more about it at: helmcode.com/deepseek-harnes…

This week we raised Deepseek limits to 2B on every user of nan.builders If you don't know about NaN, it’s a community of +500 builders sharing our GPU cluster with a monthly quota and flat fee. What’s included in NaN subscription? - 2B tokens of Deepseek v4 Flash 0731 - Unlimited tokens for Gemma4 and Qwen 3.6 - 1B MiMo v2.5 - Embeddings, rerank, kokoro, whisper, flux-2-klein All of this for your monthly €70 subscription. You can sign up for the waitlist now :) We've opened 100 more seats at nan.builders *You have a premium tier with GLM 5.2 if you need it too for €200

Token prices keep falling. But enterprise AI bills goes brrrr 📈🚀 Why is this happening? Let's dive into a quick AI economics class, because we're measuring this wrong. - A chat exchange costs a couple thousand tokens. Simple. - BUT an agentic task doesn't work like that: it plans, calls tools, retrieves context, validates its own output and retries when something fails. One task can burn hundreds of thousands of tokens before producing a result. On our platform, 41.9% of traffic is already agentic, and a single agentic session routinely consumes one to two orders of magnitude more tokens than a chat exchange. And proprietary harnesses amplify it further. Tools like Claude Code ship with heavy machinery: every turn carries system prompts, tool schemas and validation loops you don't control. Powerful, but the token meter runs on all of it. Your subscription runs out fast and your bill keeps climbing. The conclusion we'd offer any team budgeting AI for 2027: TOKEN PRICE IS THE WRONG KPI. It's the input price of a process whose consumption grows faster than the price falls. Measure cost per task completed (it's not that easy, but it's much better data) and price your infrastructure so that an agent retrying at 3am is an engineering event, not a billing one. That last part is why we price per capacity with a flat fee, not per token. Spread the word in your company: token price is not a good KPI for predictable AI billing. And stop paying per token in big 2026, go flat fee with helmcode.com

🪿 This morning we shipped our first issue of "undefined_". The Helmcode newsletter written by the team for the first 200 members. Just plain text and a few links about AI and inference. Something we talk about: - The new Deepseek Harness - New Deepseek v4 pricing - Anthropic watermarking and how it a affects to companies and open models - 2 new Helmcode events for the next months 👀 Don't miss issue 2 about our setup with Pi agent and more. Subscribe here: helmcode.com/newsletter

🐋 DeepSeek raised their API prices last week. Routers are figuring out how to pass the increase on to you. We DOUBLED down on V4 Flash 0731 instead. New monthly allowances on every enterprise plan: Starter → 5B tokens Growth → 15B tokens Scale → 35B tokens Same plans, same prices, applied automatically. Nothing to do on your side. How can both things be true at once? Because we don't resell DeepSeek's API. We serve it on our own EU infrastructure. And in case that wasn't enough: personal accounts at NaN, our AI builders community, just got their Flash usage QUADRUPLED. 500+ builders sharing a cluster paying a flat community price. If you're not part of it yet, you should come take a look: nan-builders.com That's the quiet superpower of open weights: when the developer raises API prices, everyone reselling the API inherits the hike. Everyone running the weights doesn't. 🤷‍♂️

Watermarking is more than what you have read these days about Anthropic. In fact, this affects to everyone. The EU now requires AI-generated content that reaches people to be detectable as such. We've tried to explain what really happens to every company and dev in this article: helmcode.com/posts/watermark… Every major lab signed up for watermarking. We read the technical papers AND all 50 pages of the EU Commission's guidelines and built the full explainer: - what a text watermark actually is - who's legally on the hook and for which outputs - how a company is affected - how a dev is affected - and what happens when the model is open The quick summary, in 5 points: 1/ A text watermark is invisible. It doesn't add hidden characters: it changes how the model picks between equivalent words, seeded by a private key. No quality loss, no way to spot it by reading. 2/ The law behind it: Art. 50(2) applies since August 2 (systems already on the market get until December 2). And marking without offering detection doesn't comply: it's two obligations, not one. 3/ The good part: source code is OUT of scope (including comments, SQL, YAML, configs). So is agents' chain-of-thought. So is agent-to-agent communication, in those exact words. Most real inference traffic isn't even in scope: the obligation lives where generated content meets a person. 4/ With closed models, the mark is the lab's: their key, applied globally (EU rules end up marking a Kansas user's text), and your corporate content becomes detectable by design. 5/ With open models, the mark is yours: it lives in the sampler, not the weights and the guidelines explicitly allow marking at the inference layer with your own key. In closed, the watermark is something done to you. In open, it's something you decide. Everything with link to the sources and resources: helmcode.com/posts/watermark…

The new DeepSeek Harness, explained by our team. dsh is at ~144K stars on GitHub and crushing it around the world. We've built an intensive deep dive on the harness and on the paper behind its Cordis kernel: 23 sections, from running it in 60 seconds to the formal theory, with reading routes from 5 to 40 minutes. Read the full harness explained, and how to run it: helmcode.com/deepseek-harnes… The quick summary, in 5 points: 1/ It's not a model. It's everything around one: tools, terminal, permissions, memory and the loop. The layer that turns a model that answers into an agent that works. 2/ There is no protected core. Every piece is swappable, including the one that decides what the agent does next. New behavior means adding a plugin, not forking the project. 3/ It ships with its own theory. Under the harness sits Cordis, and under Cordis an 88-page paper with formal proofs (Peking University + DeepSeek) showing a program can rewrite itself without breaking. Projects don't usually publish the math that holds them up. 4/ It's MIT and model-agnostic: it runs against DeepSeek or any OpenAI-compatible endpoint. Same harness, your choice of models and infrastructure. 5/ Same model, different harness = different agent. That's the real lesson, and it's the whole agent game right now. Tt's a developer preview and breaking changes are guaranteed. Free, EN/EN: helmcode.com/deepseek-harnes… We're not affiliated with @deepseek_ai but you can send us a bunch of teddy whales 🐳

DeepSeek open-sourced their agent harness yesterday. We spent the morning taking it apart and built an explainer site: plain language, three depth levels, from "how do I run this" down to the 88-page paper underneath. 🇺🇸🇬🇧EN: helmcode.com/deepseek-harnes… 🇪🇸ES: helmcode.com/es/deepseek-har… What actually changes: - Until now builders had two options: write an agent from scratch or accept the limits of someone else's harness. There is a third one now: compose your own from replaceable parts. - No protected core. Even the agent loop, the piece that decides what the agent does next, is a plugin with the same status as a provider adapter. Swap it and the rest keeps working. - Plugins undo themselves. Everything a plugin registers stores its inverse operation, so unloading it cleans up completely, with zero cleanup code. There is an 88-page paper with formal proofs behind that idea, published the same day. - Modes are just different harnesses over the same core. One strips down to 2 tools (terminal + editor) on purpose, to benchmark models without the harness helping. Another has the model write a small program combining tools instead of chaining 20 calls. And the part that matters most in production: observability by design. - Every session is an append-only log. Resume, fork, replay and audit all derive from that single file. - The model never reads the log directly. It reads a projection, and a runtime check guarantees everything the model sees can be reconstructed from the log. Not even an installed plugin can slip context in without a trace. - Rejected messages get logged too. You can audit what never ran. - If it cannot guarantee sandbox confinement, it refuses to run. Most software does the opposite, silently. The declared direction explains all of it: an agent that can one day modify its own harness while running. Install a tool, try it, remove it if it fails, without restarting and without breaking anything. Not a capability today. But every unusual design decision points there. All the details, here: helmcode.com/deepseek-harnes…

On Friday, July 31, our platform processed 15.9 billion tokens in a single day. A new record for us and the timing tells the story. That Friday, DeepSeek shipped V4 Flash 0731, a re-post-trained build of their efficiency flagship and we served to our 500 members community at nan.builders. With open models, the weights land, we serve them, and users are stress-testing the new build the same day The record was a release day doing what release days do on open infrastructure. The broader curve matters more than the peak: 90 days ago we were processing around 4-5B tokens daily. All platform numbers are public, updated weekly (last updated on August 9th), no signup required. Check it here: helmcode.com/opendata

Helmcode newsletter is here. There's plenty of AI content about what models can do. There's very little about what it takes to run them: the quantization that bought us 11% more throughput on the same GPU, what 495,748 classified requests reveal about real AI workloads, what the AI Act actually asks of your inference. Or every real challenge we face when serving inference at scale. That gap is why we're launching "undefined_", our newsletter. The name is on purpose. We are writing the part of this industry nobody has documented yet, written by the people operating the infrastructure, with the numbers attached. Two mails at month, no spam, unsubscribe in one click. If you build on open model, or you're deciding whether to, this is for you. First mail in 7 seven days. With a surprise event we are running :) 👉 Subscribe here: helmcode.com/newsletter