STEVEN STEINBECK
← ALL DISPATCHESDISPATCH 56

WEEK OF JAN 17 – 23, 2026

The inference-tooling land grab

Both major open-source serving stacks became venture-backed companies in the same week, a 14B model beat its base by seven points with a published recipe, and Anthropic rewrote its own constitution.

Two of the biggest open-source inference projects spun out into funded startups within about 48 hours of each other. That is the market placing a price on the serving layer specifically — not on the models.

The rest of the week was about reproducibility: an RL post-training recipe you can copy, a 3B-active MoE you can self-host, and a model constitution published under a Creative Commons license so outsiders can argue with it.

THU · Jan 22, 2026InferenceFunding

vLLM becomes Inferact at an $800M valuation

The creators of vLLM announced they have turned the project into Inferact, a venture-backed startup that raised $150M in seed funding at an $800M valuation, co-led by Andreessen Horowitz and Lightspeed. CEO Simon Mo told Bloomberg that vLLM's users include Amazon's cloud service.

It mirrors SGLang's spin-out as RadixArk the previous day.

WHY IT MATTERS

vLLM is the default serving layer for a lot of production open-weight deployments. Its two closest rivals commercializing in the same week tells you where the money thinks inference costs are going.

TechCrunchSiliconANGLE
WED · Jan 21, 2026InferenceFunding

SGLang spins out as RadixArk

TechCrunch reported that RadixArk, the company behind the SGLang serving stack, was valued at about $400M in a round led by Accel. Part of the maintainer team moved to the startup, which originated in Ion Stoica's UC Berkeley lab; former xAI engineer Ying Sheng is co-founder and CEO.

SGLang is used by companies including xAI and Cursor to accelerate serving and training.

WHY IT MATTERS

If you serve models yourself, the open-source stack you depend on now has a company behind it — worth watching for license, roadmap and support changes.

TechCrunch
WED · Jan 21, 2026AlignmentTraining

Anthropic replaces Claude's constitution and publishes it

Anthropic published a new constitution for Claude — a longer, holistic document describing the values and behavior it wants the model to embody, written primarily for Claude rather than for humans, and used directly in training.

It is released under a Creative Commons license so outsiders can reuse it and give feedback. Anthropic notes that Claude's outputs may not always adhere to the document's ideals.

WHY IT MATTERS

Anthropic says the constitution is the final authority shaping Claude's behavior and is used in training, which makes it the clearest public artifact of how a frontier lab specifies model values.

AnthropicSimon Willison
TUE · Jan 20, 2026Open weightsEfficiency

A 30B MoE that only activates 3B parameters

Zhipu AI released GLM-4.7-Flash, a 30B mixture-of-experts model with about 3B parameters active at inference, using multi-head latent attention and a 200K context window, plus a free hosted API.

It targets local coding assistance and agent workflows and progressively replaces GLM-4.5-Flash, which retires on January 30 with automatic request migration.

A 30B MoE that only activates 3B parameters
Hugging Face

WHY IT MATTERS

A 3B-active MoE with 200K context and open weights is a practical self-hosted option for coding agents, and the free API makes it cheap to benchmark against your current default.

Model cardRITS analysis
TUE · Jan 20, 2026InfrastructureEnergy

OpenAI promises its data centers will pay their own way on energy

OpenAI published a Stargate Community plan under which each site gets a locally tailored agreement so the project does not raise electricity costs for surrounding communities — from fully funded dedicated power and storage to paying for new generation and transmission.

The company says it is past halfway to its 10GW-by-2029 target, with Abilene, Texas already training and serving frontier systems and sites under development in Texas, New Mexico, Wisconsin and Michigan.

WHY IT MATTERS

Power, not GPUs, is the binding constraint on serving capacity, and utility-level commitments are how the big labs buy headroom. It shapes where cheap inference capacity ends up.

OpenAIReutersTom's Hardware
MON · Jan 19, 2026Post-trainingRL

A reproducible RL post-training run beats its base model by seven points

Nous Research released NousCoder-14B, post-trained from Qwen3-14B with GRPO/DAPO-style reinforcement learning in its Atropos RL environment on 24,000 verifiable coding problems.

It reaches 67.87% Pass@1 on LiveCodeBench v6, up 7.08 points over the 60.79% base, trained on 48 B200s over four days, with the context window extended from 32k to 40k mid-run. The company published the RL stack rather than just the weights.

A reproducible RL post-training run beats its base model by seven points
Hugging Face

WHY IT MATTERS

A small, fully documented RL post-training recipe that beats its base by seven points on a hard benchmark is exactly the thing you can copy for your own verifiable-task fine-tune.

Nous ResearchModel cardVentureBeat