Post-training is the product: Nemotron 3.5 Super VL becomes an industry overnight

Measured gains, hours after launch

The most striking detail in this week's post-training announcements is the speed. Mercor says it post-trained NVIDIA's Nemotron 3.5 Super VL ahead of its release, using small domain-specific datasets from its APEX environments — and the gains are concrete: legal full-credit rate (Pass@1) rose from 19.8% to 25.0%, APEX-SWE from 26.6% to 34.7%, and the APEX-Accounting mean rubric score from 34.6% to 41.9%. These are Mercor's own measurements, but the notable part is the data budget behind them. The company states: "With only a few hundred training tasks per domain, that's a meaningful gain". That sentence matters because it resets the assumption that specialization requires frontier-scale datasets. m:…6042

The numbers will need independent replication — they come from the vendor that sells the service — but the pattern across announcements is consistent enough to take seriously. Small, realistic professional tasks are moving open models past thresholds that matter for deployment. For a founder, the implication is that the moat question has shifted: the base model is increasingly a commodity input, and the tuning dataset is the differentiator. m:…6042

The enterprise service play

Scale AI made the enterprise framing explicit, announcing a collaboration with NVIDIA to help enterprises customize Nemotron open models as part of its GenAI Portfolio. The customer results Scale reports are the kind founders use in board decks: Skydance matched the accuracy of the frontier model it runs in production, the company says, at about 40x lower inference cost and 5x faster response times. For EY, a post-trained Nemotron 3.5 Super VL matched the accuracy of the frontier model the firm uses today at about one-eighteenth of the inference cost. A global insurance company's French-language roadside assistance copilot gained 31 percentage points of accuracy. All of these are Scale-reported customer outcomes, so treat them as vendor marketing until verified — but they set the benchmark every competing post-training vendor will now be measured against. m:…4611

The plumbing arrives on day zero

The infrastructure side moved just as fast. Nscale added Nemotron 3.5 Super VL — a 120-billion-parameter Mixture-of-Experts model with 12 billion active parameters, hybrid Mamba-Transformer architecture and KV cache quantization — to its Dedicated Inference platform for multimodal agents. Baseten put the model on its platform with day-zero training support through Baseten Loops, its managed product for supervised fine-tuning and reinforcement learning; it reports NVIDIA's numbers of a 1190 GDPval Elo rating and 47.3% on SciCode. The point both vendors are making is the same: inference and fine-tuning for the new model are available now, not in some future quarter. m:…7752 m:…2872

Prime Intellect's research with NVIDIA is the most methodologically interesting of the batch. One hundred GRPO steps on 2,500 verified tasks — about 11 hours on 128 NVIDIA GB200 GPUs — took the model from 34.0% to 57.7% on AutomationBench-AA. The team is blunt about where the difficulty was: before dedicated checks, roughly 15% of generated tasks could be gamed, and they conclude that "task quality is the bottleneck." Their solution-first approach to task generation, deriving programmatic checks from working reference solutions to prevent reward hacking, is the kind of unglamorous detail that decides whether post-training actually transfers to production work. m:…2910

Cline's contribution is a reminder that the harness matters as much as the weights. Its agent harness scored 61.8% on Terminal-Bench 2.1 with Nemotron 3.5 Super VL and thinking mode off, ahead of three other open-source harnesses at 53.9%, 51.7% and 44.9%. Cline credits agent-side choices — prompting the model to plan before implementation, and a stall detector that ends failed attempts and records structured error events. The company puts it directly: "A strong model matters. But the system wrapped around it can matter just as much." m:…0714

Tinker's year as the template

Thinking Machines Lab's one-year retrospective on Tinker reads as the thesis statement for all of the above. Tinker is a post-training API that manages infrastructure, scheduling and failure recovery so users fine-tune without owning GPU clusters; it now supports over 20 open-weight models from 4B to 1T parameters, and has added enterprise features like audit-log APIs, role-based access control and private VPC connectivity. The company's headline case study: Bridgewater's text-to-SQL model, trained on Tinker, beat Fable 5 and GPT-5.6 Sol on accuracy at 15% of their cost. That is a company-reported result, but the strategic claim is worth quoting in spirit — Thinking Machines argues the future is decentralized, with users training custom models on their own domain expertise and data rather than renting a centralized frontier model. This week, the rest of the industry effectively voted with its launches. m:…3217