Weikeng Chen | BIP-420 🐱

@l2iterative -- research partner

[Zro @zroai_]

GLM‑5.3 Flash is now live on Zro.

We’re giving it special treatment from day one, with a focus on fast, reliable serving.

More performance improvements are coming over the next few days.

[Weikeng Chen | BIP-420 🐱 @weikeng]

@zRdianjiao @OmerShlomovits

[Weikeng Chen | BIP-420 🐱 @weikeng]

@eli_crypto @bridgemindai My old friend, I am here to recommend you SuperGrok Heavy and please support @zroai_

[Weikeng Chen | BIP-420 🐱 @weikeng]

So DeepSeek did the SFT (to some extent, but overfit on the harness) and didn’t have enough time doing RL?

[AYi @AYi_AInotes]

Stanford这帮人可能找到了今年最狠的Claude、ChatGPT平替路线，我直接把手里所有闭源模型订阅全停了，这可能是今年开源AI最狠的一次逆袭。。

Stanford的人用DeepSeek V4 Flash，
同一个任务生成5个答案，
再让同一个模型自己当裁判打分，
挑最高分的那个交卷。

Terminal-Bench 2.1，
79%直接干到88%，
超过了Claude最新的Fable 5，
成本只有对方的十一分之一。

全网都在喊开源吊打闭源，
但90%的人没看懂真正的信号。

生成端早就卷到头了。
真正拉开差距的是验证。

以前大家拼模型参数，
拼训练数据，
拼基准测试高0.5。
现在发现，
同一个便宜模型多跑几次，
再自己挑出最好的，
就能干翻贵11倍的旗舰。

这就是test-time scaling的真正玩法。
模型不用做大，
推理时多花一点算力，
用生成加验证换效果。

验证本身还能继续缩放，
更细的评分标准，
多验证几轮，
把评价维度拆得更碎。

框架已经开源，
任何人都能拿DeepSeek V4 Flash直接复现。

模型是引擎，
验证才是刹车和方向盘。
引擎够猛就行，
没有刹车和方向盘谁也不敢开快。

我觉得接下来所有agent框架都会默认加上自验证这一层吧 hhh

[Weikeng Chen | BIP-420 🐱 @weikeng]

Today is a good day to run GLM-5.3 on all your repos for a security audit with OpenCode Go.

[OpenCode @opencode]

GLM-5.3 now available in Go

text · 1M context · same pricing as 5.2

[Weikeng Chen | BIP-420 🐱 @weikeng]

@0xJ_C 你不是第一个遇到这个问题的。一定不能用goal，一定不能用ultracode，一定要及时打断

[Weikeng Chen | BIP-420 🐱 @weikeng]

It varies by the protocol for PIR, but I would say that if the total amount of data a user needs to download exceeds 10MB or so then PIR would be more bandwidth saving. This then translates to the size of a block without witness data. So if the user is more than 100 blocks away, PIR starts being very useful.

[Weikeng Chen | BIP-420 🐱 @weikeng]

I have been working on this for a few months. It currently supports DPF-PIR, HarmonyPIR, OnionPIRv2, and now also ORAM with TEE client proxy. Lightning integration for paying for PIR and Nostr server discovery on the way.

[Spiral @spiral_xyz]

.@weikeng is building Private Information Retrieval (PIR) for bitcoin: query the UTXO set without revealing what addresses you are interested in. It’s a rare way to gaze into the bitcoin abyss without it gazing back.

[Weikeng Chen | BIP-420 🐱 @weikeng]

@lightcoin @spiral_xyz FHE based PIR is a bit tomato tomato

[Weikeng Chen | BIP-420 🐱 @weikeng]

@dcbuilder @thsottiaux why not sol ultracode fast (warning: significant chance of getting out of scope)

[Weikeng Chen | BIP-420 🐱 @weikeng]

@zesenhuang ChatGPT 订阅的 Luna Max 还是太慢了，感觉能让 DeepSeek 做的事情还是要给 DeepSeek

[Weikeng Chen | BIP-420 🐱 @weikeng]

Before: GPT 5.6 Sol Ultracode + /goal
(ran out of usage limit in 2 days)

Now: GPT 5.6 Sol High + “use Terra for subagents not Sol”, “let subagents do coding”, “my usage limit is tight please have mercy”, “I will notify you when CI is done, please do not loop wait”
(still at 99%)

[Weikeng Chen | BIP-420 🐱 @weikeng]

I was one of the users who sent that idea to gdb back then when he was collecting ideas :) I wonder if OpenAI can gift a banked reset for predicting what eventually goes into products. @thsottiaux

[Tibor Blaho @btibor91]

OpenAI appears to be working on a feature that would let you purchase a rate limit reset for Codex

Source: public checkout pricing config & ChatGPT web app assets

[Weikeng Chen | BIP-420 🐱 @weikeng]

My current use of Zro: (1) Kimi K3 as Sol, (2) GLM-5.2 as Terra, (3) DeepSeek v4 Flash as Luna.

The inference is fast, especially for Kimi K3. When you need independent code review, GLM-5.2 can help.

[Omer Shlomovits @OmerShlomovits]

Today we added @AMD MI325X as a new serving backend for @zroai_. Kimi K3 is the first model running on it, powered by @sgl_project.

We believe this is the first production support for the official Kimi K3 weights on CDNA3.

With 256 GB of HBM per GPU, the full model fits on a single 8×MI325X node. This is likely the lowest-cost hardware configuration capable of serving Kimi K3.

The benchmark below shows solid serving performance, even before adding any of MoonMath’s custom kernels.

Huge thanks to @digitalocean for their partnership.
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Here is everything we changed to make Kimi K3 work on CDNA3 with SGLang:

1. Rerouted MXFP4 experts to AITER’s Triton GEMMs:
AITER’s default 4-bit MoE path selects a FlyDSL kernel on gfx942, then fails during compilation. Its A4W4 kernels are gfx950-only. AITER states this in the source, and AMD’s MI35x tests note that “MXFP4 does not register on gfx942.” Without this change, K3’s MoE has no working execution path on CDNA3. We enable the Triton route automatically on gfx942.

2. Fixed an AITER GEMM crash on sliced activations:
K3 passes the tuned GEMM path a 1,536-wide slice of a 2,112-wide tensor. The launcher rejects its strides and throws; under graph capture, that becomes fatal instead of falling back. The tricky part is that PyTorch ignores the stride of size-1 dimensions, so the view reports as contiguous and .contiguous() does nothing. The fix explicitly checks for the canonical strides required by the launcher.

3. Fixed a recent regression in K3’s SiTU activation:
An August 1 upstream commit added an unguarded CUDA-only include, breaking the kernel build on every AMD GPU. Because the failure appears during graph capture, it looks like a graph issue rather than a missing kernel. We route ROCm to SGLang’s equivalent Triton implementation. Upstream, a two-line ifndef USE_ROCM guard would fix MI355X as well.

4. Selected the required graph-capture mode up front on ROCm:
Upstream replaced graph recapture with a validator that raises an error. Under speculative decoding, capturing a mode that is too weak now kills the run instead of triggering a recapture. We force the correct hidden mode before capture begins.

5. Added support for 12 MLA heads per rank:
With TP8, K3’s 96 heads become 12 heads per rank. AITER has no MLA kernel for that shape: it rejects the configuration at startup and can later call abort() from C++ without a Python stack trace. We bypass the startup assertion and zero-pad the query heads from 12 to 16 so a supported kernel can run.

6. Enabled the configuration the stack actually requires:
The working setup needs:
SGLANG_USE_AITER=1
SGLANG_AITER_K3_OPT=1
<head-padding flag>
--trust-remote-code
SGLANG_USE_AITER=1 enables the AITER MoE route. Without SGLANG_AITER_K3_OPT=1, expert weights are padded to 256 and TP8 runs out of memory during loading. Missing --trust-remote-code fails only after loading roughly 1.42 TiB of weights.--kv-cache-dtype fp8_e4m3 is not inherently required by K3 or ROCm; our current attention path requires it because it does not yet include a BF16-KV kernel.

All credit goes to the team - I am here simply to report their work 💐

[Weikeng Chen | BIP-420 🐱 @weikeng]

@meetqy 一定要用 chatgpt pro 或者 sol 规划，我是明确让 chatgpt pro 直接产生一个具体的全局方案，可以多写几个文档，要**打一个压缩包**，而且第一阶段任务要仔细拆解为一个工程的计划。

[Weikeng Chen | BIP-420 🐱 @weikeng]

@riverleaf88 @Personuo 要仔细看，建议稍微引导，ultracode 和 goal 都容易走歪，是很常见的问题。

[Weikeng Chen | BIP-420 🐱 @weikeng]

I think the good thing about dsv4flash is that it avoids the issue with Sol that Sol overworks, overthinks, and would take a long time.

[Dmitriy Ryajov @dryajov]

Ok, I was wrong. It's really good, maybe not as good as glm 5.2 but - I've seen the code and it's really good, specially on style/taste - doesn't overthink; does a good job at getting there without too much ceremony. Really good comprehension and communication. I'm impressed and for the price, unbeatable right now.

[Weikeng Chen | BIP-420 🐱 @weikeng]

I am confident that I definitely wouldn’t need this model :)

[Greg Brockman @gdb]

ten significant advances in mathematics and theoretical computer science.

solved using an internal version of Astra, our next major model, for a total cost of about $2000 at Sol API prices:

[Weikeng Chen | BIP-420 🐱 @weikeng]

Zro helps with the thinking latency in DeepSeek v4 Flash.

[Omer Shlomovits @OmerShlomovits]

The updated dsv4-Flash-0731 feels unfairly good :)
We added day-@zroai_ support serving it from our EU infra.

And, just like with Kimi K3, we cut the price 🤠