Frontline Hotspot
Frontline Hotspot

280B Total, 14B Active, Three Regions Shut Out

On November 5, 2026 Baidu released ERNIE 5.5 Flash Base and published its weights: 280B total parameters with 14B activated per token, pretrained from scratch on 33.5T tokens, 128K training context and 1M for fine-tuning and inference, 43 layers with 37 heavily compressed attention layers plus 6 MLA-DSA layers, 512 routed experts with top-10 selection and one shared expert, and FP8 plus BF16 checkpoints on Hugging Face and ModelScope. This is an end-of-pretraining Base checkpoint with no mid- or post-training, not a chat model; a post-trained version is promised without a date. The custom ERNIE 5.5 Flash Base Model License Agreement v1.0 excludes the European Union, the United Kingdom and South Korea from its territory, allows free non-commercial use, requires a prominent Powered by ERNIE notice and a one-time notice for commercial use up to 50 million dollars in related annual revenue, requires separate prior written authorization above that threshold, and bans using outputs to improve unrelated AI systems.

Published November 6, 20269 min read
<!-- ernie-55-flash-base-open-weights-hotspot | hotspot | 280B Total, 14B Active, Three Regions Shut Out -->

On November 5, 2026, the Baidu ERNIE team published the official ERNIE 5.5 Flash Base blog post and pushed FP8 and BF16 weight sets to Hugging Face and ModelScope the same day. The headline pairs two numbers with one contrast: a sparse model with 280B total parameters and just 14B activated per token gives its weights away, yet its license excludes the European Union, the United Kingdom and South Korea from the authorized territory. Weights out, door not fully open. This piece covers the event, the architecture figures and the license boundaries; it neither unpacks the thirty-row evaluation table nor explains how to run the model.

First, where this piece sits. The five-license comparison review of October 23, 2026 weighs five agreements clause by clause; this piece checks only Baidu's new license. The Kimi K3.1 Pro coverage reported a 2.8-trillion-parameter flagship delivering on an open-weights promise; this is a 280B-class pre-training-endpoint base model, a different scale and role. In this batch the base-model benchmark review owns how to read that thirty-item self-test table and choose among four bases, so this piece cites two or three numbers at most. The setup SOP handles installing, loading and serving, so this piece gives no commands. A separate line covers structRFM, an RNA foundation model resource, an unrelated subject. In short, this is the hotspot column: what happened, what the model is, where the license draws its lines.

Getting the Date and the Label Right

The date needs correcting because several versions circulate. The official ERNIE Blog post "ERNIE 5.5 Flash Base" is dated November 5, 2026, and that day the FP8 build baidu/ERNIE-5.5-Flash-Base and BF16 build baidu/ERNIE-5.5-Flash-Base-BF16 went live on Hugging Face and ModelScope; the LICENSE text states Release date: November 5, 2026 and Copyright 2026 Baidu. Chinese tech media reprinted on November 6, a reprint date, not a release date. The Hugging Face createdAt field reads October 21, reflecting a repo created before the announcement; not a release date either. Repo creation belongs to createdAt, release to the official blog and LICENSE Release date, press pickup to the media dateline. This piece uses November 5, 2026 throughout.

The label needs equal care. ERNIE 5.5 Flash Base is an open-weights model, not open source without qualification. It ships under a custom ERNIE 5.5 Flash Base Model License Agreement v1.0, and the Hugging Face model card marks license: other with license_name set to ernie-5.5-flash-base-model-license-v1.0. Open weights means the files are downloadable and usable within the agreement; it does not mean training code and data are included, nor that the grant covers the whole world. Below, those boundaries come from the LICENSE text. For now: "open weights" is accurate, while a casual "open source" erases the territory limits, the revenue threshold and the distillation ban.

280B Total, 14B Active: What a Pre-training Endpoint Looks Like

The official blog is explicit: this is a checkpoint taken at the end of pre-training, before any mid-training or post-training, a common starting point for domain adaptation, instruction tuning and reinforcement-learning research. Scale: 280B total parameters, 14B activated per token, pre-trained from scratch on 33.5T tokens, training context up to 128K, fine-tuning and inference support for 1M. The 280B figure is Baidu's official count, including the input embedding and output head and excluding the 0.36B ECHO parameters; where a different count appears, this piece does not reconcile the two.

Put 14B inside 280B and activation lands near five percent, the familiar MoE arithmetic: capacity built at 280B, a single inference pass paid at 14B. That gap is what makes releasing base weights at this scale viable, and why efficiency opens Baidu's design rationale.

Architecture, keeping to load-bearing numbers: 43 layers, one dense FFN plus 42 Multi-Router Latent MoE layers. Attention is hybrid: 37 HCA heavily compressed attention layers cache one entry per 128 tokens, while only 6 MLA-DSA layers keep token-level full-context access, so most layers see compressed history and 6 carry long-range detail. Each MoE layer has 512 routed experts with top-10 activation plus 1 shared expert, dispatching in a 2048-dimensional latent space.

Around that core sit efficiency choices: low-rank gated attention computes the gate from the low-rank query latent, cutting gate-projection parameters by 75 percent; VHA adds cross-head low-rank mixing; mHC replaces the residual stream with four parallel streams; a MultiMax output head uses a learnable piecewise temperature; ECHO reuses the backbone's final layer instead of an MTP module and doubles as a speculative-decoding draft model. All point one way, activated compute, KV cache and long-context cost as one budget.

Training uses the Muon optimizer and FP8 mixed precision. Unlike element-wise optimizers such as Adam, Muon orthogonalizes each parameter matrix's update as a whole, changing how distributed training communicates and how state spreads across GPUs. Baidu built a communication-efficient distributed Muon fusing dense and MoE synchronization into one same-rank cross-node collective, plus memory-balanced sharding. The attention kernel is in-house FlashMask-4, CuTe-DSL for SM90 and SM100, head dimensions up to 576 for qk and 512 for v. Baidu reports, as an official figure, that on SM100 it is up to 175 percent faster forward and 74 percent faster backward than FlashAttention-4 MaskMod; cite that as an official figure, not a universal result.

The weights ship as FP8 and BF16, both safetensors, mirrored on Hugging Face and ModelScope. For how much VRAM a build this size needs, or whether one machine holds it, follow the VRAM estimation and deployment SOP; this piece runs no hardware math.

It Is a Base Model, Not a Chatbot

This is the point a headline distorts. "Base" is not a version but a training stage: a pre-training-endpoint checkpoint with no mid-training, no post-training, no instruction alignment, not a chat model for direct conversation. Baidu frames it as a common starting point for domain adaptation, instruction tuning and reinforcement-learning research, a clean research baseline where post-training methods differ and a shared endpoint lets each method's contribution be compared.

That framing yields one concrete detail: tokenization must keep add_special_tokens=True, because the base model expects every input to start with a BOS token. Serving differs too: SGLang and vLLM both serve through the completions API, not a chat endpoint. Setup, launch and the first request belong to the batch's setup SOP; a hotspot piece lists no commands.

When does the post-trained version arrive? The official blog says the post-trained ERNIE 5.5 Flash "will follow" and gives no date. Under no-speculation this piece supplies no timetable. What shipped on November 5 is base weights, not a conversational product; treating it as a chatbot hits the BOS and completions details on step one.

Weights Given, So Why Not Open Source

The license is the heart of it. What follows restates the LICENSE text, no legal opinion, verified against the text fetched on November 6, 2026.

First, territory. Section 1.11 defines Territory as worldwide areas other than the European Union, the United Kingdom and South Korea, and Section 5.3 says the model, derivative models and outputs must not be used, reproduced, modified, distributed or displayed outside it. The grant is geographically bounded; those three regions fall outside. That is the opposite of open source being usable everywhere, and the source of the headline contrast.

Second, the tiers. Section 4.1 makes non-commercial use free, no notice, no revenue threshold. Section 4.2 covers commercial use where annual revenue of related products or services is at or below US$50 million: still free, but under Section 3.2 the commercial product interface must prominently display the exact wording Powered by ERNIE, and before first commercial use the operator sends a one-time use notice to Baidu_ERNIE@baidu.com. The same operator on the same model version sends one notice; once sent, qualifying use proceeds without waiting for a reply, but that notice is not the written authorization of Section 4.3. Section 4.3 covers revenue above 50 million US dollars: separate prior written authorization, applications to the same Baidu_ERNIE@baidu.com. The text stresses that an application, a notice, a pending reply or ongoing negotiation confers no above-threshold rights, and affected use lacking authorization must be suspended once the threshold is crossed. "An application is not an authorization" is a hard boundary.

Third, the distillation ban, the clause most affecting researchers. Section 5.5 prohibits using the model, any derivative model, or their outputs to improve any other AI or machine-learning system besides the model and its own derivatives. With Section 1.5, which defines derivative models to include continued training, fine-tuning, quantization, pruning, merging, adapters, and capability transfer through distillation, knowledge transfer or synthetic-data training, this draws a circle around what the weights may improve. To be precise: it does not prohibit creating or using derivative models in line with the agreement, as the Section 5.5 proviso states; it prohibits feeding them into an unrelated AI system.

Fourth, delivery scope and wrap-up. Section 1.4 states unreleased training code, data and inference code are not deliverables, and separately licensed components stay under those licenses, so "weights given" is not "full training stack given". Section 4.5 lets Baidu, at most once per calendar year per operator, or more often on reasonable suspicion of material breach, request a written compliance statement. Section 3.4 requires each recipient to assess its own use and revenue. Section 8.3 requires that after termination the operator stop affected use and delete or destroy weight copies, with legally required retention limited to that purpose. Section 3.2 attribution also applies to Section 4.3 authorized use, and Section 6.3 grants only the limited permission to use Powered by ERNIE, no other trademark rights.

Read together: territory draws a circle, revenue splits into tiers, attribution and notice are obligations, distillation has a ban, training code and data are out of the package, termination requires deleting weights. Nothing unusual for 2026 open-weights licenses, but each clause is fine-grained, down to revenue exactly equal to 50 million US dollars falling under Section 4.2. A research-only team barely meets Section 4.1; a commercial team faces a clause-by-clause checklist, not a phrase like "open source, commercially usable". How terms are interpreted in a jurisdiction, or a dispute resolved, is legal advice this piece offers none of.

Two or Three Numbers, With the Caveats Attached

Evaluation is not the focus, but ignoring it is odd, so two or three numbers with caveats attached. The official blog says it used a "unified evaluation pipeline" to compare four similarly sized base models, claiming the same harness, prompts, few-shot examples and judge model across thirty benchmarks. Any figure from that table carries two qualifiers: "Baidu unified evaluation pipeline self-test" and "no third-party reproduction available". It is Baidu's own table under its own settings, a capability profile, not an objective ranking.

Two numbers: MMLU at 93.3 and MMLU-Pro at 78.4, both official figures from that self-test. Enough. This piece reproduces no full table and aligns none of it with other leaderboards, since different setups make cross-ranking meaningless; it also avoids any cross-leaderboard absolute claim. Reading all thirty rows and choosing among four bases belongs to this batch's base-model benchmark review.

Cold Take: What the Boundaries Are Worth

To close: what landed on November 5, 2026 is base weights, 280B total, 14B activated per token, 33.5T pre-training tokens, paired with a detailed custom license. For researchers the value is a clean, un-post-trained common starting point plus a citable recipe: 43 layers of hybrid attention, 512-expert MoE, Muon with FP8, FlashMask-4. For engineering teams the value is computed with the license: the European Union, the United Kingdom and South Korea are outside the authorized territory, commercial use tiers at 50 million US dollars, Powered by ERNIE and the notice to Baidu_ERNIE@baidu.com are actions, the distillation ban constrains capability flow, and training code and data are not delivered.

Calling this "open weights" rather than "open source" is not pedantry. Open weights states a fact: downloadable, usable within terms. Open source, in industry shorthand, implies worldwide availability, free commercial use, code and data included, unrestricted derivatives, precisely the layers this v1.0 license bounds clause by clause. Conflating them costs those making decisions on the word "open source". Going forward: cite November 5, 2026 as the release date, "open weights" as the label, the LICENSE text as authority, any scores only with their self-test caveat, and "will follow" with no date for the post-trained version. What the community fine-tunes, distills or optimizes on these weights is not in the announcement; it arrives once others do the work.

Related reading

This article is AI-assisted and human-edited. Last updated: 2026-11-06

Related

Frontline Hotspot

Kimi K3.1 Pro Weights Are Out: 2.8T Parameters, 1.56TB

At midnight Beijing time on November 5, 2026, Moonshot AI published the Kimi K3.1 Pro weights on Hugging Face, honoring the promise made when the model launched on October 23. The model has 2.8 trillion total parameters with about 104B activated per token, a 1048576-token context, 69 KDA plus 24 Gated DSA attention layers, a 401M vision encoder, and MXFP4 weights with MXFP8 activations under quantization-aware training, totaling roughly 1.56TB. The vendor-reported Terminal-Bench 4.0 score is 65.7; on the third-party Agent Arena of real user sessions it ranks first among open models with a net improvement rate of 5.29 percent, ahead of MiMo-V2.6-Pro at 2.68 percent (preliminary). The Kimi K3.1 Pro License requires a separate agreement for Model-as-a-Service businesses exceeding 20 million dollars in trailing 12-month revenue and prominent naming for products above 100 million monthly users or 20 million dollars monthly revenue.

Nov 5, 20269 min read
Frontline Hotspot

Alibaba Shows Its Qwen4 Hand Early: Qwen3.8-Flash-Next Ships 125B Weights, But No License File

On August 26 Alibaba released Qwen3.8-Flash-Next: a multimodal MoE model that doubles as an early preview of the Qwen4 architecture - the same role Qwen3-Next once played for Qwen3.5. The main model is 125B parameters with an extra 51B of N-gram embeddings, activating just 6B per token; training costs about one ninth of Qwen3.7-Plus while delivering stronger coding and office performance. Four upgrades, unpacked: GDN compresses history while QSA uses a compressed indexer to pick important context at micro-block granularity; Gated Residual widens the residual stream into four branches; the N-gram embedding table can be offloaded to host memory and overlapped with compute via async prefetch; and the optimizer switches to Muon. Native context is 262,144 tokens, extensible to 1M with YaRN. The production Qwen3.8-Flash lists at \$0.16/\$0.47 per million tokens on QwenCloud (sources differ slightly; defer to the official site). The real open question is licensing: the GitHub repo ships no LICENSE file and its license field is None, the README simply points to the Hugging Face or ModelScope model page, and the community is already asking "why isn't it Apache 2.0?" - this article marks it unconfirmed, so verify the model page before any commercial use.

Aug 30, 20266 min read
Frontline Hotspot

iFlytek Spark X2.6: A 293B Flagship, but Only a 4B Is Open

iFlytek launched Spark X2.6 at its global 1024 Developer Festival in Beijing on October 22, 2026: a 293B-parameter MoE that activates 30B per token, a 1M-token context window, an end-cloud integrated design, trained entirely on domestic compute and served from a 10,000-card domestic inference cluster. The efficiency ledger, all vendor-reported: training 2000B tokens of 256K-length data on the Feixing-2 platform fell from 4.5 months to 1.2 months (a 4x gain), and the first fully domestic four-layer KV Cache architecture lifted hit rate from 63 to 95.7 percent with 4.5x inference efficiency. Self-reported scores put SWE Pro at 77.3 and SWE Multilingual at 95.0, both first, with SWE Verified at 95.4 behind only DeepSeek-V4.1-Flash 1017's 95.6 - yet there is no complete score table, no technical report, and no third-party replication. The key correction: the viral headline 'iFlytek open-sources X2.6' misreads the release. Only the on-device Spark-X2.6-4B is open under Apache 2.0 (Hugging Face under the XHToken namespace, ModelScope mirror, Ollama as SparkLLM), while the 293B flagship stays API-only. The 4B self-reports six gains including BFCL-V4 71.8 and SWE Verified 62.8. Launch pricing is 0.9 / 0.08 / 3.6 yuan per million tokens. The narrative layer predicts 60 percent local, 30 percent hybrid, 10 percent cloud workloads with routing accuracy up from 81.2 to 96.1 percent; the chairperson's 'only model trained on fully domestic compute' is a unilateral claim with no chip model disclosed. Unpublished: chip models, cluster size, total FLOPs, technical report, flagship weights.

Oct 23, 20268 min read