On October 19, 2026, the Xingchen General Intelligence Lab at China Telecom's AI technology company released and open-sourced Xing4.0-29B-Intent, repository at github.com/XingChen-AGI/Xing4.0-29B-Intent.
Repository: https://github.com/XingChen-AGI/Xing4.0-29B-Intent. It is a narrow model: a 29B MoE base activating roughly 4B parameters per token, for schema-guided intent recognition and structured information extraction. The official evaluation gives it 58 (source: the official README and the official WeChat announcement evaluation tables).
Fifty-eight points reads like another specialist-beats-the-generalists headline. Read the table column by column and the story is messier: the highest category score belongs to Gemini-3.6-Flash, not to this model. That contrast decides what the model is for, and how far the lane-change claim can be pushed.
Boundary note: every number below comes from official material — the README, the official WeChat announcement, and the official English announcement on the model page; no third-party reproduction exists. Repository counts are a snapshot from October 22, 2026. Media figures are labeled with their sourcing level.
1. Two Official Tables, Two Row Counts
The launch shipped two versions of the evaluation table. The official WeChat announcement carries 18 rows; the GitHub README keeps 11 of them and adds a Category column. README version (source: the official README evaluation table; 0-100 scale, rounded to integers, bold marks the highest per column):
| Model | Total | Category | Event Extraction | NER | Relation Extraction | Intent Understanding |
|---|---|---|---|---|---|---|
| Xing4.0-29B-Intent | 58 | 69 | 40 | 72 | 38 | 81 |
| Gemini-3.6-Flash | 49 | 71 | 30 | 71 | 29 | 64 |
| Qwen3.8-Max | 46 | 71 | 29 | 66 | 26 | 63 |
| Kimi-K3 | 45 | 71 | 28 | 68 | 25 | 61 |
| GPT-5.6-Sol | 45 | 70 | 28 | 68 | 23 | 60 |
| GLM-5.2 | 44 | 70 | 29 | 65 | 25 | 58 |
| GLM-5.3-Flash | 44 | 69 | 27 | 65 | 26 | 58 |
| DeepSeek-V4.1-Flash | 42 | 67 | 26 | 61 | 23 | 56 |
| DeepSeek-V4-Pro | 41 | 67 | 23 | 60 | 24 | 56 |
| LongCat-2.0 | 41 | 68 | 24 | 62 | 23 | 53 |
| GPT-6-Luna | 39 | 64 | 24 | 59 | 20 | 51 |
WeChat version, all 18 rows (source: the official WeChat announcement evaluation table; its columns read intent classification, entity detection, relation detection, and event extraction, matching the four README tasks):
| Model | Total |
|---|---|
| Xing4.0-29B-Intent | 58 |
| Gemini-3.6-Flash | 49 |
| Qwen3.8-Max | 46 |
| Kimi-K3 | 45 |
| GPT-5.6-Sol | 45 |
| GLM-5.2 | 44 |
| GPT-5.5 | 44 |
| GLM-5.3-Flash | 44 |
| GLM-5.1 | 42 |
| DeepSeek-V4.1-Flash | 42 |
| DeepSeek-V4-Pro | 41 |
| LongCat-2.0 | 41 |
| GPT-4.1 | 40 |
| DeepSeek-V4-Flash | 39 |
| GPT-6-Luna | 39 |
| GPT-4.1-Mini | 37 |
| Qwen3.7-Flash | 36 |
| GPT-4o-Mini | 31 |
Two counting details matter. The announcement claims a lead over the highest total of 18 mainstream general-purpose models, but row 1 of those 18 rows is the tested model itself, so only 17 are actually compared. And the full 18-row list exists in the WeChat announcement only: the official English announcement on the model page lists just five models, identical to the first five rows here. The 18-versus-11 gap is an internal difference between two official documents, not a media invention. The total column runs from 58 down to GPT-4o-Mini at 31: a 9-point lead, not a rout.
2. The Honest Contrast: The Category Crown Belongs to Someone Else
Split the gap across tasks and the composition shows: intent understanding 81 against Gemini's 64, a 17-point lead and the largest single advantage; event extraction 40 against 30; relation extraction 38 against 29; named entity recognition 72 against 71, 1 point. Four equally weighted tasks: it wins three, two decisively.
The column easiest to overlook is the category score. The top value is Gemini-3.6-Flash at 71, with Qwen3.8-Max and Kimi-K3 also at 71, while this model's 69 comes after them: on whether a request was assigned to the right category at all, it did not take first place. The official definition is narrow, evaluating only the category label with equal weight across the four tasks (source: the official README evaluation notes), so it measures whether the classification is right, not whether the extracted fields are right — precisely the ground a specialist model should own. The table notes that bold marks the highest raw score per column, judged before rounding, and identical integers do not necessarily mean a tie (source: the official README table note), so the 2-point gap between 69 and 71 is meaningful, while whether the three 71s are truly tied cannot be settled from the rounded table.
For selection, the contrast is the point: a general-purpose model suffices if the business only needs to know whether a sentence asks about an order or requests a refund; the structural advantage is where the money is if it needs the order number, amount, refund reason, and invoice status each in the correct field.
3. What This Table Can and Cannot Prove
The methodology has to be nailed down first, because 58 is easy to misread as a business metric. The test set holds 6,000 samples drawn from 48 public datasets covering the four tasks; each task score is a macro average of field-level Micro-F1, the total is an equal-weight average, and the 0-100 scores do not equal business accuracy; the category score does not mean all fields are correct, and different interfaces, prompts, and inference configurations also affect results (source: the official README evaluation notes).
The last two conditions deserve repeating. A 58 describes a relative position on an extraction-oriented evaluation, not 58 percent of production requests parsed correctly; accuracy on business vocabulary, address parsing, and order numbers is never settled by a public dataset. The official section on model information says business effectiveness has to be validated against your own data and Schema, and the model only responds quickly to key decision information, without long-form reasoning (source: the official README). These numbers are also self-evaluated: the official announcement states the evaluation was carried out by the company's own technical team and does not represent performance across all tasks or deployment scenarios (source: the official WeChat announcement); no independent third-party evaluation has been published.
4. The Second Evaluation Table and Three Latency Claims
A second table exists only in the official WeChat announcement. The first evaluation scores the category plus every field; the second scores whether the category decision is right, over 5,606 items: 3,940 single-category, 1,666 multi-category, plus 839 negatives with nothing to extract (source: the official WeChat announcement).
| Model | Total decision score |
|---|---|
| Xing4.0-29B-Intent | 70 |
| Jev | 56 |
| Jev (yes/no answering) | 53 |
| GPT-6-Luna | 65 |
Only the totals are quoted here, deliberately: both official announcements agree on every total (70, 56, 53, 65) but disagree on sub-metric names and values — WeChat calls them single-choice, multi-choice, and negative scores, the English announcement calls them Category F1, Strict all-correct, Multi-intent F1, and Negative strict, and the single-choice figures differ (source: both official announcements). Mixing the two sets would manufacture a contradiction.
Latency is where care is most needed. Official material gives qualitative claims only: the WeChat body text says the hundred-millisecond range, the official capability table says under 200ms, and the official English announcement says sub-second (source: both official announcements); the README publishes no latency metric. Media reports of latency below 150 milliseconds are a secondary claim, not an official commitment. The repository was quiet two days after launch: 17 stars, 1 fork, zero open issues, no release and no tag, created October 17 and last pushed October 19 (source: measured October 22, 2026) — too new for any third-party reproduction.
5. The Architecture: 4B Running Inside a 29B
The one-handed source for architecture facts is the config.json on the model's Hugging Face page (source: config.json fields). MoE: 64 routed experts, 4 selected per token, plus 1 shared expert, first 2 layers dense, sigmoid scoring, noaux_tc top-k. Backbone: hidden size 3584, 40 layers, 32 attention heads, 32 key-value heads, feed-forward intermediate 9216, vocabulary 131072; MLA-style low-rank attention with q at 768 and kv at 512; context 262144 (256K) with yarn rope scaling. Precision is bfloat16; the model type xing4_0 is a house architecture, so loading requires trust_remote_code=True; mHC is enabled (hc_mult 4) alongside one MTP layer. The weights ship as 41 safetensors shards, 62.43 GB (58.14 GiB), for 31,215,031,088 parameters (source: the official Hugging Face repository files and model API).
The engineering reading is straightforward: 29B of knowledge capacity paid at 4B per token. 256K of context is not a core metric for intent recognition, but it is headroom where long dialogue history has to fit, and MLA low-rank keeps the KV cache down at that length. The output protocol is restrained: a JSON array of category plus args, empty args when there is nothing to fill, multiple intents per sentence and multiple parameter instances supported. The official example turns a request about tomorrow's flight from Beijing to Shanghai into a flight-booking category with departure Beijing, destination Shanghai, and date tomorrow (source: the official README example).
The WeChat announcement also names the framework Unified Schema-driven Semantic Parsing (USP), an acronym absent from the GitHub repository, whose own wording is schema-guided (source: the official WeChat announcement and the repository description); both names stand side by side.
6. The Data, the License, and Where It Already Runs
The companion dataset Xing4.0-29B-Intent-oss was open-sourced alongside the model: about 2.75 million samples, precisely 2,753,000, drawn from 61 public datasets plus China Telecom's in-house smart-terminal assistant data, spanning 4 task types, 19 domains, and 52 languages (source: the official README companion-data section and the official data/README.md). Everything is normalized into category plus args, with Schema and sample-building scripts for mixing open and in-house data.
The label set comes with its own discrepancy. The README advertises 170+ intents and 148 slots; the released schema files measure 118 categories and 81 args for the assistant schema and 247 categories and 74 args for the phone schema, a union of 347 categories and 147 args (source: the official README and the released dataset schema files). The official side has not explained the two sets.
The license is the least frictionful item: the official text states that model weights, public data, and code including the Python SDK are all under Apache License 2.0 (source: the official README license section); the dataset page adds that the original license terms of the public source datasets also apply and should be checked before commercial use. Official quantized builds also exist: FP8 at about 33.17 GB and a single-file GGUF iq4_nl at about 17.75 GB (source: the official Hugging Face quantized repositories).
Deployment is already real. The official write-up names intent recognition and smart ticket routing on China Telecom's 10000 customer-service hotline, and spoken command parsing for smart home speakers (source: the official README and the official English announcement) — three reuses of one Schema protocol.
Separately, the lab built CMPA-SLU, a Chinese spoken language understanding benchmark, at github.com/XingChen-AGI/CMPA-SLU, whose paper has been accepted by EMNLP 2026 (source: the official README dataset section). The acceptance is confirmed, but the paper is not public yet: arXiv has no preprint and the repository citation field still reads Coming soon.
7. The Route Debate: Does a Specialist Model Really Change Lanes
In the larger picture this table is real evidence for the lane-change claim: the 58 against the 49 belongs to a specialist trained to align one sentence with pre-declared fields — a deliberate trade of breadth for precision.
But the same table supplies three counter-signals: the category score is highest for a general-purpose model; the NER lead is a single point, so the advantage concentrates in composite extraction involving event and relation structure; the total lead is 9 points with the table bottom still at 31. The conclusion is conditional: where the Schema is stable, request volume is high, and the result must be written straight into a ticket or a routing decision, the specialist payoff is structural; where the Schema changes every quarter or the answer wanted is more than one category and a few fields, a general-purpose model's flexibility cannot be replaced.
8. How This Piece Divides the Work
This article covers the release event and the route question: what happened, how the numbers were built, how selection should be reasoned about. Running the model locally, installing the SDK, and the pitfalls belong to a separate SOP article here.
One boundary worth stating. The official deployment documentation publishes no hard minimum GPU memory figure: the launch command uses two-card tensor parallelism with --tensor-parallel-size 2, and the H100 setup in the docs (80GB per card, 160GB across two) is explicitly labeled a demonstration environment rather than a requirement. An official member did give a capacity reference in the discussion area of the base repository (Xing4.0-29B-A4B, HF discussion 9): a single H100 80G can reach or come close to the full 256K context at one concurrent request, and 512K needs roughly double the KV cache headroom — with the official caveat that actual usage depends on quantization and framework overhead. The docs' own hardware statement is that Ascend 910B2 NPUs need at least two cards.
To close the loop: an official evaluation table with a 58, a category score of 69 that is not first place, a 9-point margin at the top, two official table versions that disagree on row count, latency claims that stay qualitative, and weights, data, and code all under Apache-2.0. It does not overturn general-purpose large models; it shoves them half a step aside on one narrow track, and on a track with a stable Schema that half step converts into structured results written to a database.
Sources
- Official repository README: https://github.com/XingChen-AGI/Xing4.0-29B-Intent
- Official WeChat announcement (18-row table, 5,606-item second evaluation, latency wording, USP naming): https://mp.weixin.qq.com/s?__biz=MzkwNDQ5OTU1NA==&mid=2247501059&idx=1&sn=fd0e5d8562b6e37ec1070965b332697a
- Official English announcement on the model page: https://huggingface.co/XingChen-AGI/Xing4.0-29B-Intent/discussions/1
- Hugging Face model card, config.json, and companion dataset: https://huggingface.co/XingChen-AGI/Xing4.0-29B-Intent
- CMPA-SLU benchmark repository (EMNLP 2026 acceptance): https://github.com/XingChen-AGI/CMPA-SLU
- Secondary claim, labeled in the body: media reporting of latency below 150 milliseconds