Here's a rundown of today's topics from news collected on September 17. In enterprise AI coding, Databricks reported real-world results from rolling out GPT-6 Astra to roughly 3,500 engineers, and the revelation that coding spend grew about 60% in exchange for the performance boost drew plenty of attention. On the open-model front, a 29B MoE from China Telecom's ecosystem, “Xing4.0-29B-A4B,” made headlines for being trained entirely on Huawei NPUs. In web infrastructure, Cloudflare announced a new setting that blocks AI training only. Add to that a DeepSeek engineer's essay, a 30-day real-world trial of Qwen 3.8 27B, and Xiaomi live-streaming an RL training run — a day rich with voices from the trenches.

Databricks Shares Numbers from Its Company-Wide Astra Rollout — Performance “Clearly Better,” Spending Up 60%

Databricks' Patrick Wendell reported the results of rolling out OpenAI's GPT-6 Astra to about 3,500 engineers after a pilot with roughly 200 users. According to the report, Astra “clearly” outperforms previous top-tier models on high-complexity system design and long-horizon tasks, but may not deliver major improvements on medium- and low-complexity coding.

The headline, though, is cost. Access to Astra drove total coding-related spending up by about 60%, and the company has reportedly set up a dedicated sub-budget for Astra to encourage selective use. While many benchmarks showcase cheap per-task costs, real-world large-scale deployments tend to expand usage and inflate total spend — and this is arguably the first time that dynamic has been spelled out with concrete numbers.

Around the same time, Steve Yegge, one of the most prominent champions of AI coding enthusiasm, also made news by shutting down his project “Gas Town.” Yegge admitted that even while paying hefty monthly coding-agent bills, Gas Town was the only thing he ever managed to finish with it — a story making the rounds in the community as “a bucket of cold water on the hype.”

A 29B MoE Trained Entirely on Ascend NPUs — China Telecom's “Xing4.0-29B-A4B”

The MoE model “Xing4.0-29B-A4B,” released by China Telecom Artificial Intelligence Technology (XingChen-AGI), is drawing attention on Reddit's LocalLLaMA community. It has 29B total parameters with 4B active parameters per token, natively supports a 256K context extendable to 512K, and descends from the TeleChat series.

Its biggest distinguishing feature: it is the first model of this scale trained entirely on Huawei's Ascend NPU platform with the MindSpore framework. Thanks to multi-layered optimizations — fine-grained MoE communication optimization, selective recomputation, and automatic graph and operator fusion — training throughput is reported to have improved by about 96% over the unoptimized baseline.

On performance, it scores 75.00 on SWE-bench Verified, far ahead of the comparison Gemma4-26B-A4B (53.00) and roughly on par with Qwen3.6-35B-A3B (76.00). On the engineering-focused Terminal-Bench 2.1 it scores 57.50, beating both comparators (30.00/51.50). Inference is supported on SGLang, vLLM, and KTransformers, and the model is reportedly already tuned for agent environments such as OpenCode, Claude Code, and OpenClaw.

Also making news the same day: a Mozilla analysis arguing that Chinese open-weight models have closed to within about four months of US frontier models — at drastically lower cost. While they still lag on some benchmarks, the take is that their “good enough” performance-to-price balance holds up in production use.

Cloudflare's New Setting Separates “Allow Search, Disallow AI Training”

Cloudflare has announced a new setting, “Disallow AI Training,” that lets sites block AI training by crawlers while still allowing those same crawlers to index for search. As reported by ITmedia AI+, Google, Apple, and Microsoft are expected to support it.

Under conventional operations, allowing crawling tended to mean your content also fed AI training, with no way to control the two separately. If site operators gain fine-grained control over AI training alone, this could become a new equilibrium between sites that want to preserve search traffic and AI developers. How far the big three's support extends into real-world operation is something to watch going forward.

A DeepSeek Engineer on “Letting AI Write Kernels” — Could Surpass Experts Within 6–12 Months

An essay posted on WeChat by a DeepSeek engineer is drawing a huge response, particularly on Reddit. The author claims to have written DeepSeek v4.1's main attention operator (MQA, head_dim=512), and says AI capability has moved beyond the code-assistant stage to autonomously reading CUDA/PTX/SASS, profiling per-instruction stalls, and optimizing GPU kernels. They predict that AI-written kernels could match or exceed expert humans within 6–12 months, with human roles shifting away from writing by hand toward the “pilot” side — generating and steering AI.

There's a warning too, on the education front: a concern that students who did their formative training with AI will churn out badly designed code at 10x speed without ever acquiring core skills like abstraction and system design. How this asymmetry between rising capability and skill formation reshapes engineering culture is worth watching.

Running Qwen 3.8-27B Locally for 30 Days — The Operating Costs Next Door to “Production-Ready”

A field report posted to Reddit's LocalLLaMA covers 30 days of running an Unsloth-quantized Qwen3.8-27B (UD-Q4_K_XL) locally. On a dual-GPU setup (later revealed to be an RTX 5070 Ti + RTX 4070 Super), the reported numbers are 845.1 tok/s average prompt processing, 73.8 tok/s average generation, and an MTP acceptance rate of 0.481. The verdict: production-viable for coding-agent workloads, and strong on image and UI tasks as well.

The operating-cost breakdown is just as interesting. Weaknesses cited include reasoning mode consuming up to about 50% of the context and sometimes emitting reasoning traces stretching to 60k tokens; generation speed dropping versus Qwen 3.6; and tool calls breaking down or repeating the same call beyond the 100k-token mark. Mitigations introduced include mandating sub-agents, controlling reasoning levels, and pruning broken tool-call context.

In the comments, many pointed to the effects of quantization. One report noted that on FP8, tool-call problems almost never occurred even after generating millions of tokens, with stable operation up to nearly 262k context — suggesting that Q4 quantization worsens the loop-like breakdowns and that FP8/Q8 carry a clear stability advantage.

Xiaomi Live-Streams RL Training — MiMo-V2.6's Unusual Transparency

Xiaomi's MiMo team is streaming MiMo-V2.6's reinforcement learning (RL) training on a live dashboard: multi-task RL spanning multiple agent harnesses, 1,568 prompts × 16 rollouts, fully asynchronous execution, and reward distribution based on test cases and rubrics — all shown in real time, down to batch composition and cumulative cost.

Outside estimates put the training cost at roughly $490,000 per day for a 1T-scale Pro run and about $250,000 for Flash, and reports say the dashboard's granularity drew more attention than the numbers themselves. With frontier-model training internals customarily a black box, if the practice of disclosing this much operational data spreads, RL implementation know-how could quickly level out across the field.

Wrap-up

Real-world cost figures (Databricks), Chinese open models closing the gap (Xing4.0 and Mozilla's analysis), a new web-vs-AI-training equilibrium (Cloudflare), shifting skills and roles (the DeepSeek essay), local-operation know-how (the 30-day Qwen 3.8 trial), and training transparency (MiMo). Alongside the performance-gain headlines, it felt like a day when the questions “how much does it cost?” and “who uses it, and how?” moved to the front. See you at the next digest.