
Meet Wally. Our inference stack for open frontier models, built to be the fastest place to run them. Performance snapshot: GLM-5.3 Flash: 380 tok/s. GLM-5.3 Max: 790 tok/s Qwen3.8-27B: 485 tok/s DeepSeek-V4.1 Flash: 615 tok/s 1/7

Inference research at the frontier. We make frontier open models faster and cheaper on Wally. Backed by @ycombinator W26

Meet Wally. Our inference stack for open frontier models, built to be the fastest place to run them. Performance snapshot: GLM-5.3 Flash: 380 tok/s. GLM-5.3 Max: 790 tok/s Qwen3.8-27B: 485 tok/s DeepSeek-V4.1 Flash: 615 tok/s 1/7

We built Wally Code for the future generation of builders from India. Unlimited DeepSeek V4.1 Flash inside the coding agent you already use: Claude Code, Codex, OpenCode, Hermes Agent or OpenClaw. No token meter, no 5-hour window, no weekly cap. ₹1,000 a month for students, ₹2,000 for everyone else. Billed in rupees, pay with UPI. 50% off for preorders. runwally.com/code
Introducing Wally Code: UNLIMITED coding at ₹1,000 a month for students. Same model, about 12x cheaper than @SarvamAI's API and 21x cheaper than @deepseek_ai's own API, for an agent running all month. Sarvam got picked to build India's own model and launched a per-token API. This is what India deserves, built with no grants and no subsidies.

Introducing Wally Code: UNLIMITED coding at ₹1,000 a month for students. Same model, about 12x cheaper than @SarvamAI's API and 21x cheaper than @deepseek_ai's own API, for an agent running all month. Sarvam got picked to build India's own model and launched a per-token API. This is what India deserves, built with no grants and no subsidies.
Sarvam got picked by the government in April 2025 to build India's own foundation model. They got a ₹98.68 crore GPU subsidy for 4,096 H100s and raised $234M at a $1.5B valuation in June. Now they've opened Token Factory to everyone. 8 models on it. 1 is theirs. 6 are Chinese open weights (DeepSeek, Kimi, GLM) and 1 is Google's Gemma. So India's sovereign AI lab is mostly serving Chinese models from Indian GPUs. That's a hosting business, and plenty of companies already run one. I wanted Sarvam to be the lab that ships frontier research out of India. After that much capital and subsidized compute, this is not it.

first day on a new repo i just want the tour, not the readme. gave deepseek v4.1 flash on wally karpathy's nanogpt link and got 4 stops: config, training loop, attention, forward

It's all Apache-2.0, AgentX setup included. Grab a trn2 box and you can rerun every number yourself. Blog:runanywhere.ai/blog/glm-5-3-… Code: github.com/runanywhereAI/trn…

Each GLM-5.3-Flash layer used to run as a chain of small kernels with compiler glue in between. We rewrote every layer as combination of two megakernels making it 3.66x faster

Try it at runwally.com, $5 in free credits on signup. Want it even quicker? Just add -fast: deepseek-v4.1-flash-fast

We tried this on a real restaurant site that was failing Lighthouse. Told DeepSeek V4.1 Flash on Wally to fix it and 6 seconds later it was done, we hadn't even read the first error
Every vibe coder's SEO prompt: "My vibe-coded site doesn't show up on Google. Fix it. - Switch to server-side rendering. - Create an XML sitemap. - Submit it in Search Console. - Let Googlebot in through robots.txt. - Strip out every noindex tag. - Collapse the redirect chains. - Fix every broken link. - Set canonical URLs. - Give every page a unique title and meta description. - Use one H1 per page. - Add FAQ and Article schema. - Add breadcrumb navigation. - Internally link every orphan page. - Write alt text for every image. - Compress images to WebP. - Get CLS under 0.1. - Load every page in under 2 seconds. - Rewrite anything that sounds like AI slop. - Add a real author with a real bio. - Earn backlinks from sites people trust. Page one by Monday. Make no mistakes." The joke is that this really is the full SEO checklist. Bookmark it, paste it into Claude Code and do the last two yourself.

Coding agents aren't slow because they're dumb. They're slow because every edit waits on the model. We gave OpenCode a real feature request on TodoMVC. DeepSeek V4.1 Flash on Wally shipped drag-to-reorder in about a minute, first try. Waiting is optional now.

That tiny KV cache is exactly why DeepSeek V4.1 Flash is so crazy fast to serve. Less memory per user, way more agents on every GPU. So I pointed it at the Redis codebase in OpenCode and grilled it. Exact file and line, back in seconds. Long context just got cheap.
I don't think people realize just how much of an engineering marble DeepSeek V4.1 Flash is, because guess how much RAM you need for 1M context window? For reference Llama 3.3 70B would need around 320gb of RAM. Qwen 3 8B around 144gb of RAM. DSV4.1F? Less than 0.9gb.

One dev dodged a literal supply chain attack. I wanted that for the whole registry. So I pointed DeepSeek V4.1 Flash on Wally at npm's live feed. Every new package gets read and screened the second it's published, and the sketchy ones get flagged.
Claude just saved my ass from a North Korean supply chain attack Like actually Thank you @AnthropicAI

Stack Overflow got under 1,000 questions last month. Here's what replaced it. I gave DeepSeek V4.1 Flash on Wally 3 real questions from this week. Answers came back before the page finished scrolling. RIP the green checkmark.
New record for Stack Overflow: <1000 questions posted last month

DeepSeek V4.1 Flash vs real CSSBattle targets. It looks at the picture, writes the HTML and CSS, and the first two matched in under two seconds each. Frontend devs, it's been an honor.


Security teams need models they can actually put to work. DeepSeek V4.1 Flash Uncensored is now on Wally, with Shannon, DeepSec and VVAH integrations for authorized security research.
Cheap tokens mean nothing if defenders can’t use them. Today we’re bringing DeepSeek V4.1 Flash Uncensored to Wally, with open-source security harnesses and hosted inference for authorized work. Here’s what we learned, and where I think this is going.


Everyone's adding DeepSeek V4.1 1017 Flash to their editor. I made it my autocomplete. On Wally it's so absurdly fast that a reasoning model writes whole functions as ghost text before my finger is off Tab. Copilot who?
DeepSeek V4.1 Flash is now available in Cursor! Every effort level sits on the CursorBench cost frontier: from $0.60 to $0.98 per task, no model scores higher. Turn it on in Cursor Settings > Models.