Ten days felt like a year. One lab gave away a 552-billion-parameter model for free. Another quietly shipped a model almost as good as its flagship and hid the price inside a subscription. Same country, same week of September, opposite playbooks.
The Chinese large model (大模型) field is no longer just chasing the frontier. It is arguing about what most users actually need — and pricing the answer aggressively.
DeepSeek V4.1 Flash: open weights, agent-first architecture
On September 10, 2026, DeepSeek made V4.1 Flash generally available. The headline number — 552 billion total parameters — is almost a distraction. The point is how few of them run per token.
DeepSeek describes the model as a causal encoder–decoder mixture-of-experts: roughly 8 billion parameters activate to read your input, and about 16 billion to write the output. That asymmetry is deliberate. Agent workloads spend most of their tokens re-reading context, not generating, so DeepSeek made the "reading" half cheap and the "writing" half the only part that costs much.
The model also carries a 1-million-token context and reads images as well as text natively — the separate vision variant is gone. Weights are published on Hugging Face under an MIT license, so anyone can self-host, fine-tune, or ship it commercially.
The memory trick that makes 1M tokens affordable
Long context is useless if it bankrupts you in cache. DeepSeek says V4.1 Flash holds its key-value cache at about 890 bytes per token — roughly one-quarter the HBM and one-eighth the SSD of the previous Flash generation. The model card frames it plainly: a 1M-token window now fits in under a gigabyte of cache, which is what makes serving it at scale plausible.
DeepSeek is retiring its own flagship to push this one
The most telling move is what DeepSeek did to V4-Pro. From September 14, 2026, API requests to deepseek-v4-pro are routed to V4.1 Flash at Flash prices until a V4.1-Pro arrives. In other words, DeepSeek is using a cheaper, smaller-activation model to replace the product it was charging a premium for — a pricing decision dressed as a research drop.
The new rate card (per 1 million tokens, off-peak) runs $0.15 input / $0.003 cache-hit / $0.60 output; peak hours roughly double the input and output rates. Independent benchmarking from Artificial Analysis puts V4.1 Flash at an Intelligence Index of about 39–40 under maximum-effort reasoning — strong for an open-weight model of its size, if notably verbose.
Kimi K2.8 Preview: flagship performance, closed door
One day later, on September 11, 2026, Moonshot AI fully rolled out Kimi K2.8 Preview inside Kimi Code. The model ID did not change (kimi-for-coding), so existing tools and clients needed no reconfiguration.
Moonshot positions K2.8 as close to its flagship K3 on overall quality, with more efficient thinking than K2.7 Code. It adds adjustable thinking effort — low, high, or max, defaulting to max — matching the K3 series, and opens a 1-million-token context to every membership tier. It also takes image and video input, not just text.
Here is the strategic twist: unlike K2.7 Code and K3, which shipped as open weights, K2.8 Preview is closed — available only through Moonshot's own API and apps. Moonshot has not fully explained the reversal, but the timing fits a model the company describes as "flagship-adjacent at a lower tier." For a lab that built its reputation on open releases, closing the mid-tier is a signal that the valuable model is now the one you rent, not the one you download.
Two strategies, one conclusion
Put the two side by side and a pattern emerges:
- DeepSeek bets that the winning move is open, efficient, and cheap to serve — give the weights away, win on inference cost and the agent use case.
- Kimi bets that the winning move is capturing the best experience behind a paywall — keep the strong model closed, monetize through subscriptions and API.
Both agree the old contest — "who has the biggest model" — is over. The new one is "who gives users enough model for the lowest total cost," where cost includes latency, context length, and whether you can host it yourself.
Why this matters outside China
Open weights mean V4.1 Flash can be downloaded and run anywhere, so its efficiency gains are not locked to one cloud. A clinic in São Paulo, a startup in Lagos, or a research group in Berlin can fine-tune it without sending data to a foreign server. That is a different distribution model from closed US frontier models, and it changes who gets to build on top of strong AI.
Kimi's move matters for the opposite reason: it shows that even open-weight champions are experimenting with closing the tiers that matter most commercially. The "open" label in Chinese AI is becoming conditional, not absolute.
Honest limitations
- Parameter and architecture details for DeepSeek V4.1 Flash come from DeepSeek's official release notes and model card, plus independent coverage; the "196B Engram memory" figure appears in secondary write-ups but not in the primary blog we reviewed, so we omitted it.
- Kimi K2.8 Preview's parameter count is undisclosed by Moonshot; we describe only what the company published (context, modalities, thinking tiers, positioning).
- The Intelligence Index of ~39–40 is Artificial Analysis's own metric (v4.3), not a vendor claim, and reflects only the reasoning variant under max effort.
- "Close to K3" is Moonshot's qualitative positioning; we found no head-to-head public benchmark comparing K2.8 Preview to K3 on identical tasks.
- Both releases are very recent (September 10–11, 2026); real-world latency, cost, and quality at scale are still being measured by users, not yet by long-run studies.
What readers can do now
- If you run agentic or long-context workloads, test DeepSeek V4.1 Flash's open weights on your own hardware — the 890-byte/token cache and MIT license make a self-hosted 1M-context assistant realistic for the first time at this price.
- If you are a Kimi Code user, switch daily coding to K2.8 Preview and reserve K3 for the hardest engineering tasks; the thinking-effort dial lets you trade latency for quality per call.
- If you compare models for procurement, stop ranking by parameter count — rank by cost-per-task at your real context lengths, where both of these releases change the math.
