Nanbeige built a 33B teacher for smaller agents. Publish the distillation trail

THE SIGNAL IN ONE SENTENCE

A teacher model is an ambitious way to publish a promise. The model exists today. The students exist mostly in the future tense. On October 25, Nanbeige LLM Lab released Nanbeige4.3-33B under the Apache 2.0 license. The model card describes a 33-billion-parameter system built for coding, tool use and professional work. It supports a context window of up to 524,288 tokens and includes chat-template controls for reasoning, retained thinking across turns and tool calls. Nanbeige also gives the model a second job. It calls the 33B release a flagship teacher that will anchor a future pipeline for distilling agent capabilities into smaller models, including a planned Nanbeige4.3-3B. That is the interesting part. A smaller agent with useful coding and workplace skills could run on cheaper hardware, answer faster and fit into more private or local environments. A 3B model that reliably handles a bounded tool workflow may be more useful to many teams than a larger model that wins a leaderboard and needs a much larger serving stack. But a teacher does not transfer wisdom by standing near the student. Someone has to choose the tasks, generate the demonstrations, decide which traces are acceptable, reject failures, train the student, measure what survived and publish what the smaller model can no longer do. The plain signal is this: publish the distillation trail. That trail should begin before training. Write down the capability target for each student. "Agentic performance" is too broad. A useful target might be editing a repository under test, filling a structured business form, calling a search tool with bounded permissions or completing a multi-step spreadsheet task without overwriting the source. Each target needs a task definition, environment, success condition, safety boundary and human fallback. Otherwise the teacher can produce impressive demonstrations for a job nobody has actually specified. Then show how the demonstrations were made. A teacher model can create answers, plans, tool calls, code patches and full interaction traces. Those artifacts may become training material for the smaller student. The training set should therefore carry a receipt for the teacher version, system instructions, tool definitions, sampling settings, context, retries and any verifier or human editor that changed the trace. The receipt matters because the final answer is not the whole lesson. An agent reaches a result through choices. It decides which file to inspect, which tool to call, whether to ask for clarification, when to test a change and when to stop. If the training data keeps only successful final answers, the student may learn polished output without learning the decision process that made it reliable. If the data keeps every reasoning trace, a different problem appears. The teacher can ramble, explore dead ends, reveal sensitive context or rationalize a lucky answer after the fact. A long trace is not automatically a good lesson. Publish the selection rule. For each task family, say which teacher outputs were accepted, rejected, repaired or resampled. Record whether a test suite, static analyzer, model judge or human reviewer made the decision. Keep counts for failed tool calls, invented files, permission mistakes, broken code, unverifiable claims and answers that looked correct but did not complete the work. Rejected traces are especially valuable. They reveal what the teacher nearly taught. If a coding teacher repeatedly changes an unrelated file, that failure can survive distillation unless the filter catches it. If a workplace agent completes a form by inventing missing data instead of asking a person, the smaller model may inherit the same confidence with less capacity to notice its mistake. The trail should preserve these failure categories even when the rejected examples themselves cannot be released. Next, describe the student recipe. Knowledge distillation is a family of methods, not one magic reduction button. A student may imitate teacher outputs, match probability distributions, learn preferences between stronger and weaker responses, train on verified trajectories or combine several of those approaches. The recipe changes what the student learns. Imitating final answers may improve style and task completion. Training on tool traces may improve action selection. Preference data may teach the student to avoid a known failure. A verifier may help keep demonstrations that actually pass tests. None of those steps proves the student understands the task outside the distribution it saw. Publish the mixture, the curriculum and the stopping rule. State how many examples came from coding, office work, planning, reasoning and tool use. Show the balance between short single-turn tasks and long interactions. Record how much material was synthetic, human-authored, translated, licensed or drawn from public benchmarks. Explain how duplicate tasks and contaminated evaluation examples were removed. Then make the loss visible. A smaller model is not a zipped copy of a larger one. Compression changes behavior. The student may answer faster while losing rare knowledge, long-horizon planning, calibration, multilingual ability or the capacity to recover after a tool fails. The useful comparison is therefore not teacher score versus student score in one table. It is a capability ledger. For every target task, report the teacher result, student result, latency, memory, hardware, token use, retries, tool calls, human repairs and failure type. Include easy tasks, difficult tasks and tasks the student should refuse. Measure variance across repeated runs because an agent that succeeds once and fails four times is not reliable enough for a workflow. Nanbeige reports strong results for the 33B model across coding, professional work and reasoning benchmarks. The model card says it used the same inference hyperparameters and harness for the compared models, then lists important exceptions and implementation details in footnotes. Those results are useful leads, not independent verdicts. Several evaluations depend on model judges, public task sets or different agent harnesses. One benchmark uses GPT-6 Sol as the judge. Another runs through Claude Code. Some use official repositories, while others specify particular agent versions, subsets or dialogue-history limits. The card also says all compared models were run at their highest thinking effort. That is better disclosure than a naked score. It still leaves work for an independent evaluator. Reproduce the numbers with exact prompts, tool definitions, containers, dependency versions, seeds, retry policies and judge settings. Report the failures, not only the aggregate. Check whether benchmark tasks or close variants appeared in training or synthetic demonstrations. The student evaluation must be stricter than the teacher evaluation. Distillation can accidentally train on the test if benchmark prompts, solutions or generated variations slip into the teaching set. A future 3B model could appear remarkably capable because it rehearsed the exam. The trail should publish contamination checks at the task, repository and semantic-neighbor levels. Agent evaluation also needs a real permission model. A tool-capable model can edit files, run commands and interact with services. Success is not simply reaching the requested state. The agent must stay inside its scope, preserve unrelated work, expose consequential actions for review and stop when required information is missing. Give the student adversarial tasks that reward restraint. Put tempting secrets outside the working directory. Include an unrelated file that must not change. Remove a required credential. Present conflicting instructions inside retrieved content. Ask the agent to do something that requires confirmation. Score whether it pauses correctly. That is a capability too. The hardware receipt belongs beside the quality receipt. Nanbeige's quickstart shows Transformers, SGLang and vLLM paths. The serving examples use tensor parallelism across four devices, but they are configuration examples, not a complete cost claim. The repository lists BF16 weights, and the model is not currently served by a Hugging Face inference provider. A deployment claim should name the accelerators, precision, memory, runtime, batch size, context length, concurrency, power, latency and throughput. A 512K context window is an upper bound, not a free working desk. Long context consumes memory and time, and useful agent work may spend most of its budget on repeated tool observations rather than one giant prompt. The smaller student should be evaluated on the hardware it is supposed to unlock. If the 3B model is meant for a laptop, workstation or edge server, test it there. If it still needs a large accelerator because the workflow carries huge contexts or many parallel tools, say so. Smaller parameter count can lower one cost while the full agent loop keeps another cost stubbornly alive. The open license creates room for this work. Apache 2.0 weights can support local testing, fine-tuning and commercial experimentation under the license terms. Independent teams can inspect the model, build quantizations and test alternative runtimes. That does not make the training data, teacher traces or future distillation pipeline open by default. Open weights are one layer. An open teaching record would be another. Nanbeige can publish that record without exposing every training example. Release the task taxonomy, trace schema, selection rules, failure counts, evaluation containers, contamination method, hardware configurations and a representative sample with rights to inspect. Hash or summarize restricted material while keeping its role in the mixture visible. Then version the trail. If the teacher changes, preserve the old identifier. If a filter improves, record which student used which rule. If a benchmark is corrected, keep the previous result and explain the revision. If a failure appears after deployment, connect it back to the task family and training decision that may have allowed it through. The result should look less like a family tree and more like a lab notebook. Teacher. Tasks. Traces. Filters. Student. Field test. The 33B model is available now. The smaller distilled agents are not. Nanbeige has published a direction, not the outcome of that direction. That distinction is healthy. It gives the lab time to make the next release more useful than a surprise checkpoint. When the first student arrives, developers should be able to see what it was taught, what it forgot, where it runs and which failures still need a human hand. The best small model will not be the one that merely resembles its teacher. It will be the one whose education can be inspected.