--- title: "The model you evaluated is already retired: version churn in China's foundation models" date: 2026-11-02 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-model-version-churn language: en --- # The model you evaluated is already retired: version churn in China's foundation models > DeepSeek shipped eight upgrades under two endpoint names in under two years, and Qwen and GLM move just as fast — so a large model (大模型) in production is a moving target, not a dependency you can pin once. A team ships a customer-support agent in the spring. It passes a 400-case regression suite with a 94% acceptance rate. Six weeks later the same suite scores 88%, the escalation queue doubles, and nobody has deployed anything. The only thing that changed is the other end of the API. This is the least-discussed cost of building on Chinese foundation models. The weights are downloadable and permissively licensed, which makes them feel permanent. The hosted endpoints underneath them are updated constantly, sometimes without the model name changing at all. ## Key takeaways - DeepSeek's own change log lists eight model upgrades between December 2024 and September 2026, most of them delivered under the unchanged endpoint names deepseek-chat and deepseek-reasoner. - Legacy names do not fail — they get silently re-pointed. Requests to retired model names are served by the current model and billed at its price. - Version suffixes (-0813, -1017, -2507) are the only reliable signal of what is actually running. - A "non-thinking" model can become a different product: Qwen3-235B-A22B-Instruct-2507 dropped thinking mode entirely and changed its native context to 262,144 tokens. - Teams that need reproducibility should pin dated checkpoints and self-host, or run a permanent shadow evaluation against live traffic. ## The change log is the real documentation DeepSeek's public change log reads like a release cadence for a SaaS product, not a research lab: - 26 December 2024 — deepseek-chat upgraded to DeepSeek-V3 - 21 August 2025 — DeepSeek-V3.1, adding a hybrid architecture where one model serves both thinking and non-thinking modes - 22 September 2025 — DeepSeek-V3.1-Terminus, fixing language mixing and improving code and search agents - 29 September 2025 — DeepSeek-V3.2-Exp, with sparse attention and an API price cut of more than 50% - 1 December 2025 — DeepSeek-V3.2 - 24 April 2026 — DeepSeek-V4, with the two legacy endpoint names scheduled for discontinuation three months later - 31 July 2026 — DeepSeek-V4-Flash, described as only re-post-trained from the preview on the same architecture - 10 September 2026 — DeepSeek-V4.1-Flash, with the previous V4 Flash models retired and their names temporarily routed to it Note the wording on almost every entry: "both deepseek-chat and deepseek-reasoner have been upgraded to…". The caller's code does not change; the model on the other end does. The current pricing page confirms the pattern in writing — the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but the corresponding models are retired, and requests are served by DeepSeek-V4.1-Flash-1017 at Flash prices. That is a generous migration policy and a reproducibility hazard at the same time. Nothing breaks, so nothing gets tested. ## The suffix is the specification The useful convention that has emerged is a date-coded checkpoint name. DeepSeek's price list now names DeepSeek-V4-Pro-0813 and DeepSeek-V4.1-Flash-1017. Qwen does the same with Qwen3-235B-A22B-Instruct-2507. Those suffixes carry real information: - **Qwen3-235B-A22B-Instruct-2507 supports only non-thinking mode.** The model card states there are no thinking blocks in the output and that enable_thinking=False no longer needs to be passed. Code written for the original Qwen3 — where you toggled thinking per request — now talks to a model with a different response shape. - **Its native context is 262,144 tokens**, extendable to about 1,010,000, after Qwen enabled million-token inputs for the 2507 series in August 2025. A context-limit assumption inherited from the April release is simply wrong. - **Zhipu moved the same way**: GLM-4.5 in July 2025, GLM-4.6 on 30 September 2025 with the context window widened from 128K to 200K, and GLM-4.7 in December 2025 with reported gains including Terminal-Bench 2.0 at 41% and SWE-bench at 73.8%. Three flagships in about five months. If your model inventory says "Qwen3-235B-A22B" or "DeepSeek-V4-Pro", it does not identify a model. It identifies a family. ## What disciplined teams do differently The failure mode is not that models improve; it is that improvement arrives without a change record on your side. Three habits close the gap: - **Pin the dated checkpoint in configuration, and log it with every response.** Store the model version returned by the provider alongside the output, so a regression can be traced to a specific release. - **Keep a frozen evaluation running permanently.** A small, high-signal suite — two hundred real cases, graded automatically or by a human spot check — run weekly catches silent upgrades before customers do. - **Self-host when you need bit-level reproducibility.** Open weights (开源权重) mean you can serve a fixed checkpoint on your own hardware; the hosted API means you are always on the newest one. Pick deliberately per workload. ## Where this is heading The cadence is not going to slow down; if anything, dated checkpoint suffixes suggest vendors are learning to make churn explicit rather than silent. That is a small mercy: a name you can pin is a name you can test against. According to Li An, chief scientist at BrainNet (脑机网), version churn is the tax Chinese teams pay for a release cadence measured in weeks rather than years, and most of them have not yet budgeted the evaluation work it demands. ## Honest limitations This article is built from vendors' public change logs, pricing pages and model cards, which are self-reported and may lag or omit changes that were not announced. I have counted upgrades DeepSeek chose to publish; unscheduled hotfixes, A/B tests and silent weight swaps would not appear there. Benchmark figures cited for GLM-4.7 come from Zhipu's own repository and are vendor-reported, not independently reproduced. Qwen's and Zhipu's dates are stated at month or day precision only where the vendor published them. The advice here assumes API-based deployment; teams running their own weights face a different problem — stale checkpoints and missed fixes — which this piece does not address. I have not surveyed Baidu, ByteDance, MiniMax or iFlytek release cadences, so the pace described may not be uniform across the market. ## What readers can do now - Grep your codebase for bare model family names and replace each one with a dated checkpoint string, then record that string in your logs. - Schedule a weekly run of your highest-signal test cases against every model you use in production, and alert on any score move larger than your normal variance. - Decide, for each workload, whether you want "always newest" or "never changes" — and put the ones that must never change onto self-hosted weights. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-version-churn Free to quote with attribution and a link to the original.