AI-Native Services Put a $100 Billion Market in Play

@gregisenberg sizes AI-native services at about $100 billion, priced like a service and valued like software.

Greg Isenberg calls AI-native services the best business to start right now. The work is done mostly by agents, a small team handles the rest, and the customer buys finished output: closed books, reviewed contracts. He points out that US businesses spend about $4.6 trillion a year on services, roughly six times their software spend, and that pool was locked to headcount until models got good enough.

The receipts he cites: Harvey went from roughly $100 million to $190 million in annual revenue in about five months. EvenUp sells demand letters to injury firms at around $500 each and crossed $50 million in revenue. Kick does small-business bookkeeping for $300 to $500 a month, undercuts a human bookkeeper by half, and runs gross margins above 70 percent.

His framing for accounting: a business pays about $10,000 a year for QuickBooks and about $120,000 a year for the accountant who uses it. Software fought over the smaller number for twenty years. He argues agents now do most of the accountant's work, so the finished work itself can be sold at software margins.

From @gregisenberg

Agents & Coding

Open Alternatives to Jev Keep Coming

Bespoke Labs built a 9B clone in one day; Laya runs on less than 1GB.

@DataChaz shares Nimble, @bespokelabsai's open take on Jev: a 9B model trained on 2,676 curated examples in one day that runs locally on Apple Silicon or NVIDIA GPUs. JevBench author Florian S replies that it is "quite a bit less smart than original Jev." @itsPaulAi points to Laya, an open classifier on less than 1GB that he measures 50x faster on full decisions, with a 1k context limit and weaker generalization.

From posts by @DataChaz, @itsPaulAi

Three Jev Workflows Save $8k a Month

@jasonzhou1993 automated fraud screening, signal triage and content monitoring on an open-source stack.

@jasonzhou1993 says three workflows on Jev plus Treg saved his team $8k a month: fraud detection and website signup screening, triage of buying signals, and viral content monitoring. Treg, described as OpenRouter for agent tools, is fully open-source. In replies he explains the team was losing $8k a month to credit farming before bots and fraud were blocked with the pair.

From @jasonzhou1993

Models & Hardware

Xiaomi Ships MiMo-V2.6 Pro and Flash

Pro claims parity with Claude Opus 5 on most agent benchmarks; Flash targets dual DGX Sparks.

@MiaAI_lab flags Xiaomi MiMo-V2.6-Flash at 309B parameters, 15B active, with native FP8 weights around 178 GB, sized for 2x DGX Sparks. Xiaomi MiMo's launch post says Pro "performs on par with Claude Opus 5 and GPT-5.6 Sol across most agent benchmarks." @plotarmordev calls Pro a contender for the best new model to run locally, with hardware requirements and real speed the deciding factors.

From posts by @MiaAI_lab, @plotarmordev

MiMo-V2.6-Flash-RL Benched on Two Sparks

Dual-Spark build posts 88 tok/s single stream, 156 aggregate at six streams.

@Tech2Wild ran MiMo-V2.6-Flash-RL 310B on 2x DGX Spark TP2 under vLLM with a DFlash checkpoint. Scores: 70 code, 69 math, 57 JSON, 26 prose, 85 structured tables, 88 tok/s single stream on counting. DFlash k=7 accepted up to 6.9 of 7 draft tokens per step on predictable text. The build serves a 1.87M-token fp8 KV pool at 300K context with 0.37s TTFT.

From @Tech2Wild

A 27B Model Aims at Everyday Writing

Altworld's 27B open-weights model tops its own blind writing benchmark and finishes third on public EQ-Bench 4.

Altworld released Hemmingway-1, a 27B fine-tune of Qwen3.8-27B under Apache-2.0 with 262,144-token context. On its own blind benchmark of eighty real requests, the model card says it beats GPT-6 Astra by fifty points and takes the hard-ask category 72% to 9%. It placed third on EQ-Bench 4, past GPT-5.5, Opus 4.7 and Opus 4.8. @MiaAI_lab says she will test it.

From posts by huggingface.co, @MiaAI_lab

Grok 4.7 Undercuts Fable 5.1 on Price

@MiaAI_lab measures Fable 5.1 at 400% more for input tokens and about 733% more for output.

Grok 4.7 xHigh went live and @MiaAI_lab compares it with Fable 5.1 Max: Fable 5.1 costs 400% more for input tokens and about 733% more for output tokens. She puts Grok 4.7 at 80% to 88% less with a relatively small gap in coding performance. In the replies, @vishalsingh2972 computes 8.3x cheaper output and 5x cheaper input than GPT-6 Astra or Fable 5.1.

From @MiaAI_lab

StepFun Previews a 600B Model With 27B Active

One million-token context with vision; ArtificialAnalysis rates it on par with Kimi K3.

@MiaAI_lab reports StepFun is back with a strong preview: 600B total parameters, 27B active MoE, 1M context with vision, tuned for coding, 3D and web design. ArtificialAnalysis puts it on par with Kimi K3 in intelligence, and Kimi carries over 4x as many parameters. Open weights are promised for October 15th.

From @MiaAI_lab

Qwen3.8-Flash Update Serves 8 Streams on One Spark

New build posts 117 tok/s prose and 180 tok/s code with peak memory down to 91 GiB.

@Oluwaphilemon1 runs through the Qwen3.8-Flash update for a single NVIDIA DGX Spark: 117 tok/s prose, 180 tok/s code, 8 concurrent streams, optional NVIDIA NVFP4 support, and peak memory cut from about 101 GiB to 91 GiB. The build adds a 24/7 auto-restart supervisor and cached-token reporting in every response. In the replies, @ViC305 points to his EXL3 recipe as faster per stream.

From @Oluwaphilemon1

Industry

Micro Center Cannot Keep DGX Sparks on Shelves

326 units fell to 278 in under six hours; one store went from 17 to 1.

@loktar00 reports a run on DGX Sparks at Micro Center: from 9:45 to 3:30 EST, stock went from 326 to 278, and his store fell from 17 units to 1 overnight. He notes Pro 6000 cards are near peak stock in the same window. Reply threads report local sellouts, a backorder price of $4899, and a B&H price of $5,299.

From @loktar00

Sources

Sunday, September 20, 2026 Back to the Archive Tuesday, September 22, 2026