Xing 4.0 29B-A4B Is Now Callable: China Telecom's Open MoE Leaves the Lab

The weights for Xing 4.0 29B-A4B have been downloadable from the X​ingChen-AGI organisation on Hugging Face since 16 September 2026 — that part is not news. What changed in the middle of October is narrower and more useful: the model became something you can call rather than something you can only host. A public commercial listing for Xing4.0-29B-A4B appeared on 15–16 October with a 262,144-token context window and a free promotional tier, the quantised builds caught up on 13 October, and the serving frameworks finished arriving: SGLang's support pull request went from draft to mergeable-ready, vLLM already carries the architecture, and the model card now names OpenCode, C​laude Code, OpenClaw and Hermes as adapted agent frameworks. This piece is about what actually shipped, what China Telecom's own model card claims, and which of those claims nobody outside the lab has checked yet.

Two dates, and why only one of them matters today

The repo history is unambiguous. X​ingChen-AGI/Xing4.0-29B-A4B received its initial commit on 16 September 2026, with a weights upload the same day and a README the next morning. The Apache-2.0 weights have been public for a month. A first-version article written on 17 October that announced this model as new would be wrong, and it is the exact failure the earlier coverage in this series already had to correct — the leak-era posts tracked a rename from XingChen4 to Xing4_0 across six framework pull requests before anything was downloadable.

So treat 16 September as the release date. The October development is availability: up to that point, using the model meant renting or owning the hardware. Now it does not.

What the model actually is

Xing 4.0 29B-A4B is the successor to the TeleChat line, developed by China Telecom Artificial Intelligence Technology Co., Ltd. — TeleAI — and published under Apache 2.0. It is a sparse mixture-of-experts model: 29 billion parameters in the checkpoint, roughly 4 billion read per generated token. The card describes the architecture as "mHC + MLA + MTP", and the config details that matter are these:

• Shape — 40 layers, hidden size 3,584, dense FFN intermediate 9,216, expert intermediate 1,024, vocabulary 131,072.

• Experts — 64 routed experts with 4 active per token, plus 1 always-on shared expert. That shared expert is why the active count is not 4/64ths of the total.

• Attention — MLA (multi-head latent attention), the D​eepSeek-lineage compressed-KV scheme rather than standard grouped-query attention.

• Context — 256K native, described on the card as extensible to 512K. The commercial listing quotes 262,144 positions.

• Speculative decoding — MTP, multi-token prediction, trained in rather than bolted on.

• Licence and formats — Apache 2.0, with bfloat16 safetensors, FP8, w8a8 and GGUF builds published by the same organisation.

The mHC piece is the one that is genuinely unusual. It stands for manifold-constrained hyper-connection residual streams, and it is the reason the model shows up in serving-framework code as something other than a standard transformer: the frameworks need custom kernels for it, and SGLang's implementation ships a split-K kernel specifically for the four-stream hyper-connection layout. If you are a practitioner deciding whether to bother, the honest summary is that this is a model with a real architectural bet in it, not a stock mixture-of-experts with a new name.

The numbers — and who produced them

Everything below is from China Telecom's own model card. None of it has been independently reproduced, and the card itself states the harness and sampling settings behind each row, which is more disclosure than most vendors volunteer. Read it as a vendor's claim, not a measured fact.

• SWE-bench Verified — 75.00, run under the SWE-agent harness at a 210K context window.

• Terminal-Bench 2.1 — 57.50, terminus-2 agent, 64K max tokens, 24-hour timeout, averaged over three runs.

• SWE-bench Multilingual — 66.00, same harness as Verified.

• Claw-Eval — 76.55, official harness, 256K context, three-run average.

• Tau3-Bench — 64.63, official harness, pass^1 averaged over four runs.

• DeepresearchBII — 60.80, OpenCode harness with an Exa MCP server enabled.

• AIME2026 — 90.00, five-run average. IFBench — 69.67. AA.LCR — 61.00.

Two of those are worth dwelling on. A 4B-active model claiming 75 on SWE-bench Verified and 57.5 on Terminal-Bench 2.1 is claiming to sit in territory occupied by models several times its active size, and both figures are agentic — which is to say they depend heavily on harness scaffolding, not only on the model. The card is refreshingly specific about the scaffolding, but specificity about a harness is not the same as an independent run of it.

Where it wins on its own table, and where it does not

The card benchmarks itself against Gemma4-26B-A4B and Qwen3.6-35B-A3B. On China Telecom's own numbers the picture is not a clean sweep, which is itself informative:

• Against Gemma4-26B-A4B — Xing 4.0 wins SWE-bench Verified 75.00 to 53.00, Terminal-Bench 2.1 57.50 to 30.00, SWE-bench Multilingual 66.00 to 51.00, Claw-Eval 76.55 to 71.49, Tau3-Bench 64.63 to 58.90 and DeepresearchBII 60.80 to 39.30. It loses IFBench 69.67 to 72.67 and AA.LCR 61.00 to 66.00, and trails on AIME2026 by 1.7 points.

• Against Qwen3.6-35B-A3B — Xing 4.0 wins Claw-Eval, Terminal-Bench 2.1 and DeepresearchBII, and loses SWE-bench Verified 75.00 to 76.00, Tau3-Bench 64.63 to 67.20 and AIME2026 90.00 to 92.70.

The pattern is consistent: strong on the agentic and terminal-shaped evaluations, weaker on instruction following and long-context reasoning. Those re-runs of the competitor models were performed by China Telecom, not by the vendors who built them, so the comparison is internally consistent but not independent in either direction. (An earlier article in this series covers the G​emma 4 side of that table in detail.)

One training claim worth flagging separately

Xing 4.0 was trained on Ascend NPUs with MindSpore and MindFormers, not on GPUs with PyTorch. The card claims roughly a 96% throughput improvement over its own baseline, achieved through MoE communication tuning, selective recomputation, DVM graph-operator fusion and hand-written Ascend C kernels for the mHC operators. That figure is a training-infrastructure claim about an internal baseline, which makes it essentially unverifiable from outside — there is no public reference point for "baseline" here. It is nonetheless the most interesting engineering statement in the release: a full frontier-adjacent MoE trained end to end on a non-N​VIDIA stack, with the operator work done in-house.

What it costs to run

Because the licence is Apache 2.0, the download price is zero and stays zero. Hardware is the real bill. The collection blurb claims a 4-bit quantisation at roughly 15GB of VRAM, which would fit on an RTX 3090 or 4090; that figure is a vendor-side claim and sits oddly against the 31B-parameter tensor size the repository reports for the full checkpoint, so verify it against the actual quantised file before you buy a card around it. The supported paths are SGLang, vLLM and KTransformers for inference, with LLaMA-Factory and MindFormers for fine-tuning.

For anyone who would rather not run it themselves, there is now a metered route: the commercial listing that appeared in mid-October serves the model with a 262,144-token context, a 235,929-token maximum completion, and the full sampling surface including tool calling, structured outputs and a toggleable reasoning mode. During the promotional window a free variant of the same endpoint is listed, which is the cheapest possible way to find out whether the vendor's SWE-bench row survives contact with your own repository. As always with a brand-new listing, check what you are actually being routed to before you put it in a production path — the free tier and the paid tier are frequently different deployments.

How to call it, and how we fit

To be direct about our own position: Xing 4.0 29B-A4B is not in OrcaRouter's catalogue today, so we are not a route to it and this piece will not pretend otherwise. If you want it, the options are the vendor's own API, the third-party marketplaces that have recently listed it, or your own hardware.

Where a router earns its place here is on the other side of the decision. The realistic path for most teams is to run this model in a sidecar, on a low-stakes path, for long enough to see whether its agentic strength is real on their own tasks — while production traffic keeps running on models you have already validated. OrcaRouter is one API in front of 200-plus hosted models at 0% markup, with a routing DSL and automatic failover, which means the evaluation is a config change rather than a second contract: the router keeps serving the incumbent, and you add the challenger where it has earned the right to answer. When a vendor price cut lands on a model we do carry, that price is live on our side the same day, because we pass provider list price through and add nothing to it.

What is still unverified

Four things, in order of how much they would change the verdict. First, all nine benchmark rows are vendor-reported and none has a third-party reproduction. Second, the 256K-to-512K extension is a card claim with no published recipe for the extension itself. Third, the benchmark table's competitor columns are the vendor's own re-runs, so the win margins are not comparable to the competitors' own published numbers — G​emma 4's card, for instance, does not report SWE-bench Verified at all, so there is no cross-check available. Fourth, the "93.52 on SuperCLUE" figure that has circulated alongside this release appears in a collection blurb rather than in the model card's own table, and should be treated as unsourced until it is not.

The most useful thing you can do with this model in October 2026 is the thing the announcement does not do for you: put your own hardest agentic task in front of it and see whether the terminal-bench-shaped strength survives outside a harness table. If it does, a 4B-active Apache-2.0 model that runs on one consumer card is a genuinely useful thing to own. If it does not, you have spent a weekend and one GPU. The next milestone to watch is the first independent eval run on the weights — not another framework pull request, which is now a solved problem.

None of that is hypothetical: the documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.