When I started looking into open-weight models, I saw a bigger gap. The frontier models were highly capable while open models felt like toys for only tinkerers or average code completers with limited capabilities. Recent open releases — NVIDIA Nemotron 3.5 Super VL, the Qwen4 preview models line, and what’s coming next. Previous frontier scores at a fraction of the size and cost. The exact factors can vary and it’s under debate, but one thing is sure: they also advance on techniques.
Also n-gram tables shrink the embedding weights so a large model fits in consumer RAM. On the intelligence level we see reinforcement learning.
The lever changed: from pretraining to post-training
For a long time, the honest answer to “why is model X better than model Y” was: more data, more compute. That learning is expensive, slow, and out of reach for everyone except big wealthy labs. The leaderboard was a spending chart.
RL post-training will potentially move it. Take a capable mid-size base model, put it where success is measurable. The tests pass, the tool call returns the right payload, the receipt JSON validates — and let it practice. Reasoning, tool use and instruction following sharpen hard when the training signal is did the task actually work instead of what token usually comes next. The old signal needed a huge corpus to validate. The new one needs a passing grade and a machine to check it.
Qwen’s GSPO (Group Sequence Policy Optimization) explains why this scales now. The previous generation of RL algorithms, GRPO and friends, went unstable during long runs and easily suffered irreversible model collapse. So more compute gave no more capability. GSPO grades the whole response sequence instead of token by token. This makes for a more reliable, steadier result and a model more suitable for Mixture-of-Experts training. Qwen says it “can deliver continuous performance improvement through increasing the training compute”, the property you need if RL is to be the main capability engine. It powered the current Qwen3 models (Instruct, Coder, Thinking). If you check the latest Qwen model attributes, you can see: it is already a Qwen4 preview.
Punching up: what the numbers actually say
Nemotron 3.5 Super VL is the cleanest recent demo. A 120B MoE activating only 12B per token — cheap to serve — that beats models two steps above it on agent benchmarks:
| Model | Hermes Bench | LangChain Evals |
|---|---|---|
| Nemotron 3 Ultra (its own “big brother”) | 56.33% | 84.3% |
| Nemotron 3.5 Super VL | 63.44% | 87.05% |
A distilled model outscoring the model it was distilled from: the student passed the teacher by training harder on the job, not by being born smarter. It clears 600 tokens/second on agentic tasks on a single GPU and scores 80.8% on CharXiv chart reasoning (84% with code interpretation).
The same punch-up shows up on the Qwen side of the table. Qwen3.8-Flash-Next — the first public model on the architecture previewing Qwen4 — is a 125B MoE that activates 6B parameters per token, and its post-trained checkpoints beat Claude Opus 4.6 (Max) on the agentic benchmarks: SWE-bench Pro 62.5 vs 53.4, JobBench 55.7 vs 36.6, IFBench 81.3 vs 62.5. On CharXiv chart reasoning it reaches 90.6, where Nemotron Super VL scored 84% with code interpretation.
These are only vendor-reported numbers on each lab’s own harness, but still probably not overly far from the truth.
The mechanism repeats across labs. Post-training stacks SFT on agent trajectories, then multi-teacher on-policy distillation — teachers specialized in video reasoning, vision tool use and alignment score the student’s own generations — and RL in agentic environments on top. The base model is already smart and the RL-heavy post-training makes it even better.
How close is “close”? Measuring the gap honestly
“Punching up” is not just a feeling — the numbers exist. A May 2026 LessWrong study (17 benchmarks, ~110 datapoints) measured how long it takes an open-weight model to cross thresholds the frontier already crossed:
We find that, as of today, on private benchmarks, where the data is not publicly accessible, open models are roughly 8-10 months behind the closed frontier, while for public benchmarks the gap is roughly 4-6 months. We also find that the gap was smallest around the time of DeepSeek R1, in Jan 2025, and since then the gap has been growing.
Still: the distance is measured in months, not years.
Why doesn’t that gap invalidate the playbook below? It measures generic capability, and a real workflow is not generic. Once a model is RL’d on your own environments, you compare a model that knows your tools and policies against one that doesn’t. A few weeks of post-training can make those months irrelevant for that task, codebase and workflow.
The playbook: this is now your move too
The real advantage here: they come with open weights, open datasets and open training recipes. The lever that compressed the frontier gap from years to months is on this side of the table too.
NVIDIA shipped Super VL with roughly 40 vision-language datasets plus agentic tool-use, finance reasoning and instruction data on Hugging Face, and NeMo pipelines for SFT, LoRA and RL against your own environments. The early adopters are reaping the production wins:
- Scale AI post-trained Super VL for automated vendor credit processing at Skydance: 40x lower inference cost, 5x lower latency, beating the frontier models previously on the job.
- Ramp converts receipt images to structured JSON with it: expecting up to 50% higher throughput and 10x lower latency.
- AI21 and Alice post-trained it for threat intelligence: +41% accuracy over the baselines.
The pattern: take a strong open mid-size model, RL it against a replica of your workflow — your tools, your policies, your document formats — and a 12B-active model beats a frontier model that knows nothing about the business.
"Frontier level" is becoming a per-workflow statement, not a per-model one.
Qwen’s Qwen-Scope ships trained Sparse Autoencoders for Qwen3/3.5 that expose internal features: find the activation pattern behind endless repetitive generation, amplify it during RL sampling so those bad cases actually get generated and can be optimized away. Targeted synthesis from rarely activated features reportedly improves data efficiency roughly 15x. A training tool, not an audit report.
What to do with this
Anything with a clear success signal — validated output format, passable tests, checkable facts, replayable environments — is a candidate for a post-trained open model at a fraction of the cost, wherever data governance demands it.
However, keeping up with open models can still be costly as the model cycles are quite rapid. If you wanna move with the frontiers, you have to re-train and customize the newly introduced models again.
The models stopped being the scarce resource; sometimes the amount of releases feels like even noise.
However, workflows, environments and success criteria are the scarce input, and RL turns them into capability.
The lever is public. The models are shinier and shinier.
Sources
- NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning — NVIDIA Technical Blog, Oct 2026
- Group Sequence Policy Optimization (GSPO) — Qwen blog
- Qwen3.8-Flash-Next: A New Architecture, Towards Ultimate Cost-Efficiency — Qwen blog, Aug 2026
- Qwen-Scope: Decoding Intelligence, Unleashing Potential — Qwen blog
- How far behind are open models? — Håvard Tveit Ihle, LessWrong, May 2026 (data & code)