Let there be tokens. Open-weight inference for token-hungry workloads: long-running agents, evals and batch at a fraction of the cost. app.doubleword.ai
Your token bill is too high.
Models are getting better, agents are getting more ambitious, but the bill keeps getting worse.
We don't think developers should have to ration intelligence. One customer cut their bill 4x on the same model and workload.
Let there be tokens.
You can now connect Doubleword straight from @opencode
Type /connect in opencode, search for Doubleword and paste your API key. Our models then show up in the model picker, ready to code with.
Pick from open models like Kimi K3.1 Pro, GLM-5.3 and DeepSeek V4.1 Flash, and get coding in your terminal or the desktop app.
I’ve been playing around with @opencode this week and asked GLM-5.3 (@Doubleword_ ) to make me an Ulam spiral 😵💫 using a Julia script.
In 1963 the mathematician Stanisław Ulam was bored in a talk and started doodling. He wrote the numbers 1, 2, 3… in a square spiral going outwards, then circled the primes.
Primes are meant to look pretty random, so you’d expect it to be very noisy, but they regularly appeared in diagonal lines.
I think that is pretty cool because each line traces a formula that happens to be unusually good at producing primes. There’s also a 100-year-old conjecture that predicts exactly how good, though it still hasn’t been proven.
I ran it for primes up to 10 million. The Julia script checked every number, placed it on the spiral and coloured each prime by type i.e. twin primes (two apart, like 11 and 13), cousin primes (four apart, like 7 and 11), and so on.
We recently announced that Doubleword is now a provider on LangWatch.
To demonstrate, Rogério Chaves put Doubleword to work on a freight inbox triage agent.
Across 50 hand-labelled emails, DeepSeek V4.1 Flash on Doubleword matched GPT-6.1 Terra on quality at 82% less per email, and 90% less on Doubleword Batch.
👇 Check out the full blog at: langwatch.ai/blog/doubleword…
Doubleword is now a provider on LangWatch
Test, evaluate and trace agents at scale with a swift and setup.
Drop a Doubleword API key once, and choose a combination of open models like Kimi K3.1 Pro, GLM 5.3 and DeepSeek V4.1 Flash for your workloads and evaluators.
Thanks to @ManoukDrai47041, @_rchaves_, and the LangWatch team for the integration!
We have constructed the pareto frontier for pokemon showdown performance...
Notable findings:
🐳DSV4.1 Flash is on the pareto frontier, Kimi is not very good, GLM is really bad
🤠GPT-6 models are way better than Opus
👨👩👧👦Models predict the behaviour of other models in their own families better than models from other families
🧵
We have constructed the pareto frontier for pokemon showdown performance...
Notable findings:
🐳DSV4.1 Flash is on the pareto frontier, Kimi is not very good, GLM is really bad
🤠GPT-6 models are way better than Opus
👨👩👧👦Models predict the behaviour of other models in their own families better than models from other families
🧵
🚀 Now live on Doubleword: Day 0 support for Kimi K3.1 Pro.
@Kimi_Moonshot's latest flagship model is now available on Doubleword, built for complex coding, agentic workflows and long-context reasoning.
⚡️ Model highlights:
1M-token context window
2.8T parameters, with 104B active (MoE)
Multimodal support for text, image and video inputs
Strong performance in agentic coding, full-project development and complex reasoning
Start building → app.doubleword.ai
Tomorrow @MeryemArik9 joins Sam Roddick at @Deloitte's Financial Services AI Expo to discuss sovereign AI and scaling AI with control over infrastructure, data and cost.
🕝 2:30–3:00 PM
📍 London
Congratulations to @radostin_st, @finn_fergus and their collaborators who published last week their work 'parallel restore-optimized snapshot compression for cost-efficient inference serving'
we use this kind of research from our inference lab to cut cold start times when we autoscale models so we can start generating tokens as soon as the GPU is up!
Check it out here: radostin.io/files/stoyanov-c…
If you want work on making the next token cheaper to produce than the last please reach out! We're hiring!
Congratulations to @radostin_st, @finn_fergus and their collaborators who published last week their work 'parallel restore-optimized snapshot compression for cost-efficient inference serving'
we use this kind of research from our inference lab to cut cold start times when we autoscale models so we can start generating tokens as soon as the GPU is up!
Check it out here: radostin.io/files/stoyanov-c…
If you want work on making the next token cheaper to produce than the last please reach out! We're hiring!
Next stop: @dawncapital's Product & Tech Summit 🚀
@MeryemArik9 is there this afternoon.
If you're building AI products and thinking about inference costs, reliability or scaling up, come say hello 👋
What does sovereign AI actually mean for UK startups? 🇬🇧
Tomorrow, @MeryemArik9 joins @Twinpathvc AI Breakfast Club to discuss infrastructure, open models and the future of UK sovereign AI.
☕ 8:30am, Shoreditch
luma.com/06bpyjsa