The Week That Was: Open-Weight AI Giants, Robotaxi Launches, and a CRISPR Milestone

AI Models: The Open-Weight Arms Race Heats Up

If there is one theme that dominated the model-release cycle this week, it is that open weight + tight pricing + broad distribution is now the default playbook for smaller labs trying to close the gap with closed APIs. Two releases in particular—Mistral Large 3 and MiniMax M3—show how far the ecosystem has come in just a few months.

Mistral 3: A 675B MoE Under Apache 2.0

Mistral AI shipped the full Mistral 3 family on June 18. The headliner is Mistral Large 3, a sparse mixture-of-experts model with 675 billion total parameters but only 41 billion active parameters per token—a density that makes inference surprisingly cheap. Mistral says Large 3 ranks #2 among open-source non-reasoning models on LMArena and #6 among all open-source models, achieving parity with the best instruction-tuned alternatives on general prompts. To put those rankings in context, LMArena is a crowdsourced evaluation platform that measures real chat quality across thousands of user prompts, and breaking into the top ten means the model holds its own against GPT-4.1-class systems in blind comparisons.

The pricing is aggressive: $0.50 per million input tokens and $$1.50 per million output tokens. That puts it in the same ballpark as several smaller API-only models and dramatically below the per-token cost of training and operating a 675B-parameter model from scratch. Alongside the flagship, Mistral released three edge models—Ministral 3B, 8B, and 14B—each with base, instruct, and reasoning variants and native image understanding. Mistral claims the 14B reasoning variant hits 85% on AIME 2025, a math-heavy benchmark that separates surface-level pattern matching from genuine problem-solving.

Deployment is a big part of the pitch. Mistral partnered with NVIDIA, vLLM, and Red Hat. A checkpoint in NVFP4 format lets Large 3 run on a single 8×A100 or 8×H100 node. NVIDIA also optimized Blackwell attention and MoE kernels for GB200 NVL72 systems, while DGX Spark, RTX PCs, and Jetson devices handle the edge models. The models are already live on Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, Modal, IBM WatsonX, OpenRouter, Fireworks, Unsloth AI, and Together AI.

What makes this release notable is not just the size but the license. Apache 2.0 means enterprises can fine-tune, redistribute, and deploy without the compliance gymnastics that come with non-commercial or source-available licenses. For companies that have been waiting for a permissively licensed model that actually competes with GPT-4-class systems, this is the strongest signal yet. The commercial implications are straightforward: you can bundle Large 3 inside a product, ship it to customers, and even sell access to a fine-tuned variant without asking permission.

MiniMax M3: 428B Parameters, 1M Context, Open Weights Delivered

Shanghai-based MiniMax followed up its June 1 launch by publishing the actual open weights and an arXiv technical report for MiniMax M3—a move that fulfilled a public commitment and earned immediate credibility from the open-source community. The model is a 428B-parameter mixture-of-experts system with a 1 million token context window that natively handles text, image, and video inputs. Only 23 billion parameters activate per token, making it computationally lean alongside being feature-rich.

On benchmarks, M3 scores 59% on SWE-Bench Pro, a coding-evaluation suite that has become a de-facto standard for agentic software engineering. That score places it among the strongest open-weight models for real-world code generation, not just synthetic benchmark performance. Independent verification by Artificial Analysis—the benchmarking firm that caught several inflated claims earlier this year—confirmed the sparse-attention architecture after scrutinizing both the weights and the paper.

The architecture itself is worth understanding. MiniMax built a custom sparse attention mechanism—documented in a separate arXiv paper (arXiv:2606.13392)—that allows the model to process long sequences without the quadratic memory cost of standard attention. For developers, this means you can pass an entire repository, a legal contract, or a long research transcript into the model and get coherent, grounded responses. For enterprises, it opens up use cases that were previously impractical: contract review across hundreds of pages, codebase-level refactoring assistants, and legal-due-diligence automation.

The timing matters. Mistral and MiniMax released within days of each other, both under permissive licenses, both optimized for efficient serving, and both targeting the same enterprise developer audience. Meanwhile, Kimi K2.7 Code from Moonshot AI entered the ring as an open-weight coding specialist with a 256K context window, and JetBrains shipped Mellum2, a 12B MoE designed specifically for code completion and chat inside IDEs. Cohere also launched North Mini Code, its first developer-facing model, signaling that every major AI vendor now sees coding as a vertical worth vertical-specific models.

The practical upshot: if you are building an AI-powered product today, the cost of switching from a closed proprietary API to an open-weight alternative has dropped sharply. The hardware requirements are real, but the tooling—vLLM, llama.cpp, NVIDIA NIM, Ollama—has matured to the point where a single node can serve models that would have required a cluster two years ago. Startups and research groups that could not afford hosted API bills at scale now have a credible path to self-hosting.

Autonomous Driving: Vegas, the Ring, and a Recall

On the vehicle-autonomy front, the week had everything: a dramatic demo, a regulatory filing, a landmark recall, and a major industry partnership. The sector is clearly moving from controlled-test-phase stories to real-world infrastructure and regulatory milestones.

Tesla Robotaxi Files for Vegas Launch

Tesla applied for an autonomous vehicle network permit with the Nevada Transportation Authority, a filing that would allow the company to run commercial ride-hailing in the Las Vegas Valley. The application was made public last week and will now go through the NTA’s staff-review process before the full Authority deliberates in a General Session. According to state spokeswoman Teri Williams, this review could take weeks or months.

Meanwhile, Tesla has already leased a 37,000-square-foot industrial building at 6170 Mohawk St. in southwest Las Vegas. Clark County records show a $3.1 million retrofit plan that includes eight superchargers and six car lifts—clear evidence of fleet-maintenance infrastructure. Tesla Robotaxi also holds two business licenses from Clark County: one tied to the Mohawk building (listed as auto wash and detailing) issued on June 2, and another tied to a building at 3338 Fremont St. (listed as administrative and support services) issued on February 10. The corporate licensee is Tesla Robotaxi LLC, headquartered in Austin, Texas.

Tesla is actively hiring—a nighttime supervisor and a nighttime fleet support specialist, both based in Las Vegas. The company previously received approval to test autonomous vehicles in Nevada beginning last September, but the new permit would move it from testing to paid passenger service. If approved, Vegas could become the first major U.S. market where Tesla Robotaxi operates as a commercial business rather than a research program. The city’s tourism-heavy economy, predictable weather, and investor-friendly regulatory posture make it an obvious first choice.

Xiaomi YU7 GT Autonomously Laps the Nürburgring

On the same day, Xiaomi announced that its YU7 GT completed the first autonomous lap of the Nürburgring Nordschleife in 10 minutes, 29.483 seconds. For context, a professional driver in the same car set a production-SUV record of 7 minutes, 22.755 seconds back in May. The autonomous lap is roughly three minutes slower, but the headline is not the speed—it is the technical demonstration. A production car navigating one of the world’s most demanding racetracks without a safety driver is a credible stress test for perception, planning, and control systems at high speed. The Nürburgring’s blind crests, 73 bends, and unpredictable surfaces leave virtually no margin for error in sensor fusion or trajectory planning.

Xiaomi described the lap as accomplished using its in-house intelligent-driving stack, though the exact sensor suite and compute platform were not disclosed. The company framed the achievement as "a new starting point rather than an end point," suggesting more aggressive autonomous performance is on the roadmap. For an EV company that entered the car business less than four years ago, the capability gap between Xiaomi’s software stack and legacy automakers is closing faster than most analysts expected. The YU7 GT itself was developed in close collaboration with Xiaomi’s smartphone and AI labs, leveraging the company’s existing expertise in computer vision and real-time inference.

Robotaxis Scale Up—and One Gets Recalled

On the partnership front, Stellantis, Wayve, and Uber announced a joint push to deploy Level 4 robotaxis globally. The deal combines Stellantis’ manufacturing scale—the parent of Jeep, Ram, Peugeot, and Fiat—with Wayve’s vision-first AI and Uber’s rider network and dispatch infrastructure. It is one of the broadest cross-industry robotaxi coalitions to date, targeting fleet deployments across North America and Europe in the coming years. Wayve, based in London, has been building a sensor-light autonomous stack that relies primarily on cameras rather than the expensive lidar arrays used by most competitors; Stellantis’ vehicle platforms give the partnership immediate hardware reach.

Meanwhile, Baidu Apollo Go received a Level 4 autonomous driving permit in Switzerland under the AmiGo brand, running in partnership with Swiss Post’s PostBus. That makes it one of the first Chinese autonomous-driving companies to win a major European commercial permit, and it positions Baidu to expand into other dense European cities with favorable regulatory frameworks. The Swiss permission covers designated routes in selected urban and suburban zones, not yet unrestricted city-wide coverage—but it is a significant legal precedent.

Not every story was forward-looking. Waymo recalled 3,871 of its fifth-generation robotaxis after NHTSA identified a defect in the autonomous driving software: the hazard-detection system failed to recognize work zones on highways, a failure mode that could lead to unsafe lane-keeping or emergency-braking behavior. The recall reflects the reality that autonomous-driving software, like any safety-critical system, must handle edge cases at the boundary of its training distribution. Waymo is deploying an over-the-air software fix, but the incident illustrates that scaling robotaxis means scaling safety engineering at the same pace.

Biotech: CRISPR and Prime Editing Hit Clinical Stride

While AI and autonomous vehicles grab most of the headlines, biotech is quietly delivering a series of breakthroughs that will define medicine for decades. This week produced three distinct milestones in gene editing, each addressing a different layer of the same challenge: how to edit DNA inside a living human body safely, efficiently, and precisely.

Phase 3 CRISPR Trial for Hereditary Angioedema Completed

Researchers from Amsterdam UMC, collaborating with multiple hospitals, successfully completed the first Phase 3 trial of an in vivo CRISPR gene therapy—this time for hereditary angioedema, a rare but debilitating genetic disorder that causes unpredictable, sometimes life-threatening swelling attacks. The therapy is delivered directly inside the patient’s body rather than by extracting and editing cells outside the body, a route known as in vivo editing. Completing Phase 3 means the treatment has crossed the highest clinical-evidence threshold short of regulatory approval, and the data will now be reviewed by the European Medicines Agency.

This is not the same CRISPR therapy that sickle-cell patients received recently; hereditary angioedema involves a different target gene and a different delivery strategy. What unites them is the demonstration that in vivo CRISPR is no longer experimental—it is a viable clinical modality. The Amsterdam team used a lipid-nanoparticle delivery vehicle to encapsulate the CRISPR components and direct them to the liver, where the disease-causing mutation is expressed. The trial design included a randomized placebo arm, and the treatment met its primary endpoint with statistical significance.

RUBY Trial: CRISPR Cures Sickle Cell in 96% of Patients

Separately, the RUBY trial—published in the New England Journal of Medicine—reported that CRISPR-Cas12a gene editing achieved a functional cure for sickle cell disease in 96% of patients. Sickle cell disease affects millions of people worldwide, primarily of African and Mediterranean descent, and until recently the only cure was a matched bone-marrow transplant, which carries severe risks of graft-versus-host disease and is unavailable to most patients. A one-time gene-edit therapy that eliminates the underlying mutation is a genuinely historic result.

The RUBY trial used CRISPR-Cas12a rather than the more familiar Cas9 nuclease, a choice that may reduce off-target editing risk. Patients were treated with a single infusion of edited hematopoietic stem cells, which then repopulated the bone marrow with healthy red blood cells. The 96% functional cure rate—defined as freedom from vaso-occlusive crises for at least 12 months—is among the highest ever reported for any sickle-cell therapy in any trial, conventional or experimental.

Prime Editing Goes In Vivo With Lipid Nanoparticles

The third biotech story comes from the David Liu lab at the Broad Institute, where researchers published three studies improving prime editing—in Nature Nanotechnology and two issues of Nature Biotechnology. Prime editing is more versatile than conventional CRISPR because it can perform precise substitutions, small insertions, and deletions at any targeted site, potentially repairing the vast majority of known disease-causing mutations. That versatility has made it one of the most anticipated tools in genomic medicine since its first description in 2019.

The bottleneck until now has been delivery: prime editors are large proteins, and getting them into cells in the body efficiently and safely has been difficult. Viral vectors such as AAVs have limited packaging capacity, can trigger immune responses that prevent redosing, and may result in prolonged transgene expression that increases opportunities for off-target editing. Liu’s team optimized the key components of the prime editing system—including the prime editing guide RNA (pegRNA) and the prime editor protein itself—and adapted the system for delivery via lipid nanoparticles (LNPs). LNPs are entirely synthetic, customizable delivery vehicles already used in approved mRNA vaccines and gene-therapy products, so the regulatory path for LNP-based prime editing is relatively well understood.

The result is higher editing efficiency and stronger potency when delivered directly into animal models, a critical step toward in vivo human therapies. The team reported successful corrections for diseases including chronic granulomatous disease, phenylketonuria, and alpha-1 antitrypsin deficiency in preclinical models. Combined with the CRISPR Phase 3 data, the message is consistent: 2026 is shaping up to be the year gene editing moved from research labs to clinical reality, and the combination of in vivo delivery with prime editing could expand that reality to hundreds of genetic diseases that were previously undruggable.

Why These Three Stories Belong on Your Radar

AI models, autonomous vehicles, and gene editing are often discussed as separate tracks, but they share a common pattern right now: permissionless innovation is happening faster than regulation can keep up, and the companies that ship usable products first are building durable advantages in data, distribution, and trust.

The Open-Weight Tipping Point

Mistral and MiniMax are not just releasing models; they are releasing entire deployment ecosystems. An Apache 2.0 weight, a quantized checkpoint, and a serving integration with vLLM or NVIDIA NIM means that a startup with a few GPUs can now compete with a cloud vendor on inference cost. This is why OpenAI, Anthropic, and Google are under pressure to justify their premium API prices—not because closed models are worse, but because the open alternative is good enough for most workloads and infinitely more flexible. When you self-host, there is no third-party uptime risk, no API rate-limit negotiation, and no vendor lock-in on prompts or fine-tunes.

The Robotaxi Infrastructure Race

Las Vegas is becoming the de-facto proving ground for commercial autonomous ride-hailing in the U.S., and Tesla is not the only player investing there. Baidu is running AmiGo in Switzerland, Wayve and Stellantis are targeting global scale through Uber, and Waymo is expanding despite the recall. The companies that win will be the ones that master fleet logistics, regulatory relationships, and public trust—not just the underlying model. The hardware retrofit costs alone (Tesla’s $3.1 million facility, eight superchargers, six lifts) suggest the barrier to entry is rising, and that favors incumbents who can amortize capital over large vehicle fleets.

CRISPR as a Standard-of-Care Candidate

The RUBY trial and the Amsterdam Phase 3 completion are data points in a larger trend: regulators are now comfortable evaluating one-time gene-edit therapies. The FDA approved the first CRISPR-based sickle-cell therapy in 2023, and now additional indications are in late-stage trials. If prime editing with LNP delivery reaches the clinic within the next two years—as Broad Institute collaborators expect—it could expand the addressable market from rare blood disorders to metabolic and neurological diseases currently considered untreatable. The economics are also shifting: as manufacturing scales and competition increases, the price of gene-edit therapies could drop from the current two-million-dollar range toward something accessible in middle-income countries.

What to Watch Next

On the AI side, the most consequential near-term event is whether Mistral ships its planned reasoning version of Large 3 and how it stacks up against o3-class systems on coding and math benchmarks. MiniMax will need to demonstrate real-world adoption of M3 beyond benchmarks, especially given the 1M context window’s steep memory requirements. Expect more model makers to publish dataset composition documentation as government scrutiny of training data increases and copyright litigation continues.

For autonomous vehicles, the key milestones are Tesla’s Nevada permit approval timeline and Waymo’s post-recall safety audit. Xiaomi’s next move—whether it translates track performance into consumer ADAS features in upcoming YU7 variants—will be closely watched in China’s already crowded EV market, where BYD, Nio, and XPeng are all racing to match its software capability. The Stellantis/Wayve/Uber partnership will likely reveal its first deployment city by early 2027, and that city will set the template for how legacy automakers compete against pure-play tech entrants.

In biotech, watch for prime editing LNPs entering Phase 1 trials—Broad Institute collaborators have hinted at 2027 targets for PKD and other liver-targeted genetic diseases. The CRISPR-based sickle-cell approval has already changed reimbursement conversations; insurers are now negotiating coverage for a $2 million one-time therapy, a process that will set the template for every gene-edit treatment that follows. The pricing outcome will determine whether these cures become global public-health tools or remain limited to wealthy health systems.

The Bottom Line

None of this week’s stories are political, and none are speculative. They are deliverables: models you can download and fine-tune, robotaxis you could theoretically hail when regulations allow, and gene therapies that have already passed human trials at the highest standard of evidence. That is what makes the current tech moment different from the hype cycles of the past. The technology is no longer "coming soon." In several domains, it is arriving—and the teams that are shipping today will define the next decade of how we build software, move around cities, and treat genetic disease.