Nanbeige4.3-33B

# Fix list: Nanbeige4.3-33B

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/nanbeige-4-3-33b, the October 2026 research run, assessed 25 October 2026. Grade C, 61 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on Nanbeige4.3-33B: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Reach, 55 out of 100, up to 17.3 more on the total

Why it scored 55: Memory earns 10, since the only build from Nanbeige or a runtime's library is 66.26 GB of BF16, about 80 GB with a fifth added, which fits the reference machine but not 64 GB. Runtimes earn 20 from released vLLM, SGLang, Transformers and llama.cpp, which all run the Llama architecture, and Ollama and LM Studio have no build. Builds earn 0, as Nanbeige publishes no quantised build. Terms earn 15 for Apache 2.0 with nothing added, though only the metadata names it. Download earns 10, with no gate. Hosted earns 0, as no provider or Nanbeige API serves it.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reach):

Whether an agent can run it, and where.

- 0 to 30, memory. The smallest build from the publisher or a runtime's own library that keeps 4-bit precision or better, at its file size plus a fifth for the runtime and a working context. 30 for 16 GB or less, 25 for 24 GB, 20 for 32 GB, 15 for 64 GB, 10 for what fits our reference machine (the GPU share of 128 GB of unified memory), 5 for 256 GB, 0 above.
- 0 to 25, runtimes. 5 for each of llama.cpp, Ollama, vLLM, SGLang, LM Studio and Hugging Face Transformers that runs the architecture in a released version, from the runtime's own code, library or documentation. A publisher's fork or an unreleased branch doesn't count. 25 at most.
- 0 to 15, builds from the publisher. 15 for quantised builds at two precisions or more (FP8, 4-bit, GGUF), 10 for one, 0 when only other people's quantisations exist.
- 0 to 15, terms of use. 15 for an OSI-approved licence with nothing added, 10 for a custom licence that lets anyone use it commercially for free below a threshold or under a use policy, 5 for research or non-commercial use only, 0 for no licence or unclear terms.
- 0 to 10, download. 10 when the files download with no account and no terms to accept, 5 for a click-through that needs an account, 0 when a person approves each request.
- 5, hosted too. The publisher's own API serves it, or at least two inference providers do, so an agent without the hardware can still call it.

When our reference-machine runs start, speed on that machine (tokens per second and time to first token at the build that fits) joins Reach.

## 2. Risk, 60 out of 100, up to 12.3 more on the total

Why it scored 60: File safety earns 25, safetensors only and no repository code, so the card's trust_remote_code has nothing to run. Provenance earns 15, the lab's own organisation as its arXiv report links it, with a hash for every file. Safety evaluation earns 0 and agent guidance 0, since neither card says anything on either. Licence risk keeps all 15 under Apache 2.0. The security programme earns 5 for BOSS Zhipin's security response centre, which isn't linked to the model.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-risk):

What could go wrong when you run it, and what the publisher has done about it. A higher score means less risk.

- 0 to 25, file safety. 25 for weights in safetensors or GGUF only, 15 when pickle files (.bin, .pt, .pth) sit beside them, 0 for pickle only. 10 off when the documented way to load it or to build its prompts runs code from the repository (`trust_remote_code`, or the publisher's scripts in place of a template).
- 0 to 15, provenance. 10 when the files come from the publisher's own organisation, the one its site or docs link to, and 5 for a checksum on every file.
- 0 to 20, safety evaluation. 20 when results are published and someone other than the publisher ran them or can rerun them, 10 for the publisher's own results, 0 for none.
- 0 to 15, guidance for agent use. 15 for documented prompt-injection guidance or an instruction hierarchy the model was trained to follow, 10 for a statement of what its trained refusals cover and the safeguards to add around it, 5 for general cautions, 0 for none.
- 0 to 15, licence risk. 15 to start, 5 off for each of these, a use policy or field-of-use limits beyond the law, terms the licensor can change or revoke, a revenue or user threshold, and a licence that changed without a dated note or whose link is broken. Not below 0.
- 0 to 10, a security programme. 5 for a way to report a flaw (security.txt, a disclosure policy or a security contact), 5 for advisories handled in public.

When our reference-machine runs start, measured robustness (prompt injection with AgentDojo, over-refusal with XSTest) joins Risk.

## 3. Documentation and upkeep, 56 out of 100, up to 6.8 more on the total

Why it scored 56: The model card earns 11, with intended use, parameters, context and languages but no limitations, architecture or cut-off. Training data earns 0, as nothing describes it. Last change earns 20 for 25 October 2026. Responsiveness earns 10, a Nanbeige member answering four of ten discussions on release day. Version policy earns 5 for a first release, with a config change made and reverted under bare commit titles. File health earns 10, the files and the card agreeing.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-upkeep):

- 0 to 20, the model card. Intended use, limitations, architecture, parameters, context length, languages and knowledge cut-off. 20 for all seven, 3 off for each that's missing.
- 0 to 20, training data. 20 for the data published or listed with sources and sizes, 15 for a data card that names the sources or the mix, the cut-off and the filtering, 10 for a general description with a cut-off, 5 without one, 0 for none.
- 0 to 20, the last change to the files, the template or the reference code. 20 within 30 days, 15 within 90, 10 within 180, 5 within a year, 0 older.
- 0 to 20, responsiveness. The publisher's replies to the 40 most recent questions and reports in the repository's discussions or issues. 20 for most, 10 for some, 5 for a few, 0 for none.
- 0 to 10, version policy. 10 for earlier versions kept in their own repositories with a dated note when one is replaced, 5 for earlier versions kept without notes or a first release, 0 when files were replaced in place without a note.
- 0 to 10, file health. 10 when config, tokenizer, template and card agree and no report of a broken file waits without a reply, 5 when one does, 0 when the files contradict the card.

## 4. Agent interface, 85 out of 100, up to 2.3 more on the total

Why it scored 85: The chat template earns 20, Jinja in tokenizer_config.json. The tool-calling format earns 25, with the XML call format in the template and the qwen3_coder parser named for vLLM and SGLang. Trained tool calling earns 20, several calls per turn in the template. Structured output earns 0, as nothing mentions it. Reasoning control earns 10, thinking switched per request with the qwen3 reasoning parser named. Context earns 10 for 512K set in the config with no scaling.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-interface):

What a model and a developer need to talk to it. Speech, image and other models with no chat interface read the tool-calling and structured-output lines for what the model takes and returns (voices, timestamps, streaming, typed outputs), and the note says how.

- 0 to 20, the chat template. 20 for a template in the repository (Jinja, in the tokenizer config or its own file), 10 when the prompt format is only given as code you run.
- 0 to 25, the tool-calling format. 25 when the publisher documents how tools are declared and how calls come back, with the parser to use in a named runtime, 10 when only the template or a runtime's own recipe says it.
- 0 to 20, tool calling the model was trained for. 20 with several calls in one turn, 10 for one call at a time.
- 0 to 15, structured output. 15 when JSON or JSON Schema output is documented for the model with how to ask for it, 5 when the card claims it without saying how.
- 10, reasoning control. Thinking that can be turned off or budgeted per request, with where the reasoning text comes back.
- 0 to 10, context. 10 for 128K tokens or more trained, 5 for 32K to 127K. Context reached only by scaling the position encoding counts one band down.

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- Capability is computed later. No independent result was found on 25 October 2026 (Artificial Analysis, Vals AI, Scale SEAL, Terminal-Bench, SWE-bench, BFCL); LMArena answered 403. So Capability stays pending and the total is capped at 61
- The repositories have no LICENSE file. Apache 2.0 rests on the metadata tag, scored by its terms under the calibration. Worth asking Nanbeige to add the file
- Who makes it is established from the lab's arXiv report (Boss Zhipin), the template's default system prompt and the card's kanzhun.com contact. That Kanzhun Limited is the legal entity wasn't confirmed first-hand, as ir.zhipin.com refused our reader
- Context earns 10 on the config's 524,288 tokens with no scaling. Nanbeige doesn't say the trained length; if 512K was reached by raising rope_theta without long training, a stricter reading gives 5 and Interface 80
- Tool-calling format earns 25 on the template's XML format and the named qwen3_coder parser. The card shows no returned call; reading that as half the requirement gives 10 and Interface 70
- Reasoning control earns 10 on enable_thinking and the named qwen3 parser. The card doesn't say in words where the reasoning comes back; a stricter reading gives 0
- Version policy 5. The config change and revert on 24 and 25 October carry bare commit titles; reading them as files replaced without a note gives 0 and Upkeep 51
- Responsiveness counted on ten discussions all opened on release day; recheck at the next pass
- The weights date is 24 October 2026 by the upload commit. The Hub doesn't show when the repositories turned public; the first discussion is from 12:36 UTC on 25 October
- unchecked, the registration date of zhipin.com

## Weaknesses

- No independent benchmark result yet; every agentic number on the card is Nanbeige's own run
- BF16 only from Nanbeige, about 80 GB to load, and no Ollama or LM Studio build
- No LICENSE file in either repository; Apache 2.0 rests on the metadata tag
- No safety evaluation, agent-use guidance, training data description or knowledge cut-off
- No hosted API or inference provider serves it

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- Serve with `vllm serve Nanbeige/Nanbeige4.3-33B --enable-auto-tool-choice --reasoning-parser qwen3 --tool-call-parser qwen3_coder`, as the card says, or SGLang with the same two parsers
- Keep `tool_call_format` at its default `xml`, since qwen3_coder parses that format; the `json` option writes Hermes-style JSON calls, which need a different parser such as vLLM's hermes
- Pass `chat_template_kwargs` `{"preserve_thinking": true}` for multi-turn tool use and coding, and `{"enable_thinking": false}` to skip reasoning, as the card recommends
- Use temperature 1.0, top_p 0.95 and up to 65,536 new tokens for agent work, and temperature 0.6 for reasoning and chat, per the card
- Send your own system prompt, since without one the template injects a Chinese identity line, or a Chinese tool-expert prompt when tools are passed
- Send one system message at the start only, as the template raises an error for a later one
- In plain Transformers pass `use_cache=True` to generate, since config.json and generation_config.json set it false; trust_remote_code isn't needed

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.