←
Nan Jiang89 posts
Nan Jiang avatar

Nan Jiang

@nanjiangwill

building fun stuff @modal

◷ Joined January 2018▧ 5 Photos and videos
Nan Jiang avatar
Nan Jiang@nanjiangwill·Nov 2

THE fastest weight sync(and you dont need RDMA) + it is open source 🤗

Large-scale RL has a problem: weight updates from trainers to samplers are slow and expensive We’re sharing weight teleportation: faster weight updates than existing intra-cluster techniques with internet speeds, enabling samplers anywhere (maybe even space!) Weight teleportation is so efficient (<0.23s sampler pause time, ~3s non-blocking compression + prepare), the bottleneck ends up being the GPU hardware’s decompression ASIC We unlock these gains by using GPUs for massively parallel compression of model weight changes, engineering a heterogeneous decompression and patching pipeline, along with other optimizations and overlaps. These techniques are co-designed with hardware-native features in NVIDIA’s Blackwell generation of chips We’ve open-sourced weight teleportation, so everyone can use it today

Liangsheng Yin avatar
⟳ Nan Jiang retweeted
Liangsheng Yin@lsyincs·Oct 24

Yes, comparing with their published numbers is actually unfair to DeepSeek, so we just did an apples-to-apples comparison on the same machines, which is more persuasive, professional, and rigorous.

Worth noting the 1.04x is against the stronger baseline. Your measured DeepInfer on 4x B300 already runs about 9% above DeepSeek's published curve at 4K and 8K (roughly 65K vs 60K and 78K vs 72K tok/s/GPU), so against the README numbers that 4K point would read closer to 1.13x

CLS avatar
⟳ Nan Jiang retweeted
CLS@ChengleiSi·Oct 21

I had tons of fun working on pretraining (& automating the pretraining) of Flint! It's a really good local model and a small preview of what we are cooking!

Introducing Flint, @Recursive_SI’s first model: a fast local coding model built with our automated research system, the Eureka Machine. • Best-in-class fast local coding • Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark • Built from scratch in under 4 months Private, unmetered intelligence on your own machine. Open weights available now on @huggingface, run it locally on your favorite framework or try it for free on @opencode or @pidotdev.

2:41
gongy avatar
⟳ Nan Jiang retweeted
gongy@_gongy·Oct 13

Everybody buying compute for RL needs to stop immediately and think about this result. Big homogenous clusters are a sledgehammer for what is essentially an inference-shaped nail. Fantastic work from @nanjiangwill.

Modal@modalOct 13

Training large models with RL typically runs on a fixed homogeneous cluster of the latest accelerators for numerical stability and faster throughput. This throws away inference optimization work done for other vendors and generations, limits the capacity you can utilize, and is unnecessarily expensive. So we trained a model on agentic coding tasks using fully asynchronous RL and mixed-precision, mixed-accelerator samplers. With off-policy correction, training stayed stable through convergence, on par with homogeneous samplers, at a lower cost.

0:12
Akshat Bubna avatar
⟳ Nan Jiang retweeted
Akshat Bubna@akshat_b·Oct 13

RL is so much more elastic if the FLOPs for rollouts can come from any hardware. If only there were a product that's good for spinning up lots of different kinds of hardware elastically...

Modal@modalOct 13

Training large models with RL typically runs on a fixed homogeneous cluster of the latest accelerators for numerical stability and faster throughput. This throws away inference optimization work done for other vendors and generations, limits the capacity you can utilize, and is unnecessarily expensive. So we trained a model on agentic coding tasks using fully asynchronous RL and mixed-precision, mixed-accelerator samplers. With off-policy correction, training stayed stable through convergence, on par with homogeneous samplers, at a lower cost.

0:12
Martin Marek avatar
⟳ Nan Jiang retweeted
Martin Marek@mrtnm·Oct 13

A cool application of score centering: running RL on heterogeneous hardware It uses the Miles implementation of score centering, which supports top-p sampling. Applying score centering to top-p sampling is a generalization of the same "mask replay" correction that Cognition and DeepSeek use for top-p training (under zero TIM, these corrections are equivalent). Another benefit is that under top-p sampling, it is feasible to run exact score centering (which is infeasible with full sampling).

Modal@modalOct 13

Training large models with RL typically runs on a fixed homogeneous cluster of the latest accelerators for numerical stability and faster throughput. This throws away inference optimization work done for other vendors and generations, limits the capacity you can utilize, and is unnecessarily expensive. So we trained a model on agentic coding tasks using fully asynchronous RL and mixed-precision, mixed-accelerator samplers. With off-policy correction, training stayed stable through convergence, on par with homogeneous samplers, at a lower cost.

0:12
Nan Jiang avatar
Nan Jiang@nanjiangwill·Oct 13

Spent the past three weeks on this idea. A sampler on a different precision or accelerator is just off-policy, like stale samples in fully async RL, and off-policy correction doesn't care why the sampler differs from the trainer. Meanwhile, inference keeps diversifying across precisions(MXFP8, MXFP4, NVFP4) and accelerators outside RL. So why not use all of that compute for RL rollout? The sampler and the trainer are two systems built for different purposes. Hope this opens up new possibilities for RL algorithms and systems, and leads to cleaner system designs. Enjoy the blog, and stay tuned for the next one on our infra :)

Modal@modalOct 13

Training large models with RL typically runs on a fixed homogeneous cluster of the latest accelerators for numerical stability and faster throughput. This throws away inference optimization work done for other vendors and generations, limits the capacity you can utilize, and is unnecessarily expensive. So we trained a model on agentic coding tasks using fully asynchronous RL and mixed-precision, mixed-accelerator samplers. With off-policy correction, training stayed stable through convergence, on par with homogeneous samplers, at a lower cost.

0:12
Nan Jiang avatar
Nan Jiang@nanjiangwill·Oct 9

so good

slime v0.4.0 is probably our biggest architectural upgrade since the initial release! A lot has changed under the hood to support RL at a much larger scale, while maintaining full backward compatibility with existing custom components.

Sasha Rush avatar
⟳ Nan Jiang retweeted
Sasha Rush@srush_nlp·Sep 15

Lean Verified Transformers (srush.github.io/lean-transfo…) In which we prove a bunch of Transformer invariants from scratch in Lean, and speculate about how hard it would be to do that for the rest of the world's code.

Post attachment
Liam Fedus avatar
⟳ Nan Jiang retweeted
Liam Fedus@LiamFedus·Sep 15

We built high-throughput materials labs in Menlo Park to create a loop between experiments and models. The labs generate fresh data, the models learn from it, and then help us decide what to try next. Using only 1,300 H200s, plus months of our experimental data, we mid-trained and RL’d an open-source model to surpass GPT-6 Astra on our analysis benchmark. We call it Neon. This is real footage from our lab. We’re focusing first on hard problems in materials science, including superconductors, magnets, and semiconductor materials. Read our blog posts below.

0:16
RadixArk avatar
⟳ Nan Jiang retweeted
RadixArk@radixark·Sep 9

The full technical report for Miles is now available. Miles is built for production-level post-training, with stability, efficiency, and flexibility at its core. The report covers its system design and how it works in practice.

Post attachment

Today we're launching Miles v0.1, an open-source RL framework for LLMs and multimodal models. RL training is easy to start and hard to debug. Miles helps you ensure your run is correct, use hardware efficiently, and keep RL running at scale. Over the past 9 months, 72 contributors have landed 1,326 commits, 85 GPU E2E CI tests, battle-testing Miles on frontier open models like Kimi K3, DeepSeek V4, Qwen 3.8, GLM 5.2, Inkling, MiniMax H3, etc. Miles powers frontier-model development and production RL workloads at @humansand, @periodiclabs, @modal, @DecagonAI, @Eigent_AI, @nebiusai, @IBM and more, on both @NVIDIAAI and @AIatAMD hardware. Here is what we built, and why teams picked Miles🧵

0:03
Shannon Shen avatar
⟳ Nan Jiang retweeted
Shannon Shen@shannonzshen·Sep 3

We share our early research on building Software World - a "GitHub" run by agents. We deploy agents for Python packages in a dependency chain, and each agent is tasked with collaborating with others and optimizing the package it owns.

0:13
Nan Jiang avatar
Nan Jiang@nanjiangwill·Aug 18

huge

Today we're launching Miles v0.1, an open-source RL framework for LLMs and multimodal models. RL training is easy to start and hard to debug. Miles helps you ensure your run is correct, use hardware efficiently, and keep RL running at scale. Over the past 9 months, 72 contributors have landed 1,326 commits, 85 GPU E2E CI tests, battle-testing Miles on frontier open models like Kimi K3, DeepSeek V4, Qwen 3.8, GLM 5.2, Inkling, MiniMax H3, etc. Miles powers frontier-model development and production RL workloads at @humansand, @periodiclabs, @modal, @DecagonAI, @Eigent_AI, @nebiusai, @IBM and more, on both @NVIDIAAI and @AIatAMD hardware. Here is what we built, and why teams picked Miles🧵

0:03
Alex Wa avatar
⟳ Nan Jiang retweeted
Alex Wa@_djdumpling·Aug 13

🧵I post-trained Qwen3-4B-Instruct to play MegaGem, @JaneStreetGroup's new auction game. It ranks #1 on 3-player MegaGem, above models like GPT-5.5, Claude Opus 4.8, and Gemini 3 Flash, but statistically unresolved with Gemini 3.1 Pro. However, self-play RL wasn’t what got it there.

Post attachment