AI builders are attacking different bottlenecks at once. Contrastive-LM released CLM-8B, ZeroDrift made Anchor 3.0 generally available, and Alibaba unveiled the Zhenwu V900 chip on September 23, 2026. Together, the announcements point to an industry chasing faster agents, tighter controls, and far more computing power.
CLM-8B trades generation for speed
CLM-8B is the first open model in a new class called Contrastive Language Models, or CLMs. It does not generate text; instead, it scores candidate actions against the current state and returns probabilities, a narrower job that avoids making a language model narrate every decision before taking it.
The CLM GitHub repo serves CLM-8B through a TypeSafe-compatible API. Its 75 MB model trains a state encoder and an action encoder with a bidirectional InfoNCE loss, scores each candidate through the dot product of those embeddings, and applies a softmax to produce the answer distribution.
Contrastive-LM trained the system on about 60 million Nemotron DQA question-answer pairs, followed with about 30 million synthetic hard negatives and about 1 million agent trajectories. Pre-training alone reached 52.1% top-1 accuracy on about 100,000 questions, mid-training lifted that figure to 69.2%, and training on hard negatives peaked at 62.4%.
The speed claim is substantial: CLM runs 13 times faster than Jev with about 1,000 candidates. Jev entered limited early access on 15 September 2026, giving the comparison a clear rival rather than a conveniently empty scoreboard.
On the T-Rex game, CLM-8B posted 16.5 milliseconds of latency and a 5/5 success rate, compared with Jev at 149.8 milliseconds and 5/5. Tool calling on BFCL v4 took 76.8 milliseconds for CLM-8B and 125.5 milliseconds for Jev, with success rates of 95.2% and 99.2%.
WikiRacing showed the sharper tradeoff: CLM-8B took 79.8 milliseconds and completed 26/30 tasks, while Jev took 225 milliseconds and completed 30/30. Super Mario produced 33.5 milliseconds and 5/5 for CLM-8B, versus 132.6 milliseconds and 5/5 for Jev.
That pattern continues on verifier benchmarks, where CLM runs 4.1 to 5.7 times faster than Jev but trails on accuracy. Fine-tuned CLM scored 73.7% on DeepSWE against Jev’s 81.6%, and 84.0% on Terminal-Bench 2.1 against 87.6%.
Compliance models and colossal chips
ZeroDrift announced the general availability of Anchor 3.0 on September 23, 2026. It is the first family of small language models built to enforce regulatory and company rules on AI-generated communications, catching 95.5% of violations in a benchmark and outperforming GPT-5.6 Sol on recall, precision, and F1.
Anchor 3.0 comes in Mini, flagship, and Max versions. Mini has 9 billion parameters with 4 billion active, the flagship also has 9 billion parameters with 4 billion active, and Max has 27 billion parameters; depending on the version, ZeroDrift post-trained them from Gemma E4B or Qwen3.8-27B.
ZeroDrift says Anchor runs up to 34 times faster and up to 12 times cheaper than frontier models. Kumesh Aroomoogan described the rationale plainly: “A frontier model is too slow and too expensive to do that on every message. Anchor is built for exactly this job. Developers keep building with the models they choose, and Anchor makes sure what their agents say is compliant.”
ZeroDrift is backed by $10 million from a16z speedrun, PitchDrive Ventures, and other investors. Developers can sign up for Anchor 3.0, while Cristian Felix said, “Strong governance is essential for deploying AI at scale, giving financial institutions the confidence and control they need to move from experimentation into production.”
Alibaba’s announcement addresses the opposite constraint: not fewer computations, but vastly more of them. Eddie Wu said the Zhenwu V900 is “the most powerful AI chip in China today,” and Alibaba says the chip delivers three times the performance of the previous Zhenwu M890.
Alibaba also plans to train a model with five to 10 trillion parameters. Its latest Qwen3.8-Max has 2.4 trillion parameters, while Moonshot’s Kimi K3, released in July, has 2.8 trillion; the parameter race has stopped pretending that restraint is part of the roadmap.
The company plans to expand its data centers beyond 20 gigawatts of computing by 2032. SpaceX had about 1.4 gigawatts of AI computing capacity in mid-2026 and aims to exceed 10 gigawatts in 2027.
Neil Shah said using extra computing power helps China stay strong locally amid chip limits. Parv Sharma said the AI and chipmaking gap between China and the U.S. depends on advancing chip foundries, placing Alibaba’s hardware push inside a larger contest involving Nvidia and Huawei.
Wu compared AI growth with the industrial revolution and predicted machines will produce more than 1,000 times as much thinking as all of humanity combined, up from less than 3% currently. The three announcements make the strategy plain: use compact models for decisions, dedicated models for compliance, and enormous infrastructure for everything else.
Based on
- Contrastive-LM Releases CLM-8B: An Open System One Model That Scores Agent Actions Up to 9× Faster Than Jev — marktechpost.com
- Alibaba unveils what it calls China’s most powerful chip and ambitious AI model plans — nbcnews.com
- ZeroDrift Launches Anchor 3.0, the First Family of Models Built for Enforcement Runtime of AI Agent Communications | Markets Insider — markets.businessinsider.com
