DeepSeek V4.1-Flash: MIT, 552B, a quarter of the KV-cache, and the end of the Pro model behind the deepseek-flash endpoint as of September 14

DeepSeek V4.1-Flash: MIT, 552B, a quarter of the KV-cache, and the end of the Pro model behind the endpoint as of September 14

🔎 Three days that reshuffle the open-model market

On September 10, 2026, DeepSeek published the V4.1-Flash weights on Hugging Face, under the MIT license, on the very day of launch. No API exclusivity window, no few-week wait: the 510 GB FP8 checkpoint, split into 48 shards, was downloadable immediately. According to Baseten, this is DeepSeek's third open-weight "flash" release of the year, in a week where Z.ai and Alibaba also released their flash models.

But the real story isn't the timeline. It's the target. V4.1-Flash isn't a model designed to top academic benchmark leaderboards: it's a model designed for agent workloads, where 98% of billed tokens are context re-reading, not generation.

And there's the operational urgency. On September 14 at 12:00 PM Beijing time, deepseek-v4-pro is being retired: all requests will silently route to V4.1-Flash. If you have hardcoded model IDs in your pipelines, you have a retest to run before that date. One developer has already measured 30% of usable context lost due to context tables frozen on the old IDs. We'll walk you through all of it.

Key takeaways

• V4.1-Flash was released on September 10, 2026: 552B-parameter backbone MoE + 196B for the Engram memory module, 1M-token context, MIT weights on Hugging Face (~510 GB FP8 in 48 shards).
• Asymmetric Causal Encoder-Decoder architecture: 8B active parameters in prefill, 16B in decode. The first major architecture designed agent-first.
• KV-cache reduced to 890 bytes per token: a 1M context fits in ~0.93 GB, versus ~2.15 GB naively in FP16 — four times less than V4-Flash.
• Disruptive pricing: $0.003/1M on off-peak cache-hit. Over a real agent tracked for 30 days, $12.78 on DeepSeek vs $64.00 on GLM-5.3-Flash at identical traffic — a 5x gap caused by a single line in the pricing table.
• On September 14: deepseek-v4-pro routes to V4.1-Flash, and the old IDs deepseek-v4-flash and -vision-exp also resolve to the new model. Re-test your pipelines.

Recommended tools

An asymmetric architecture: why 8B for reading and 16B for writing

Because reading and generating don't carry the same computational cost — and DeepSeek is the first major lab to structure its entire architecture around this asymmetry.

The model is built on a Causal Encoder-Decoder (CED) that splits the 40 layers into two blocks: 20 causal encoder layers, and 20 decoder layers whose KV cache is projected directly from the encoder's output. Concretely, prefill only runs the encoder (8B active parameters), while decode runs at 16B. The logic, as explained by Baseten: generating demands more than reading, and agent workloads produce far more prefill tokens than decode tokens.

Under the hood, other components are worth a closer look, detailed in Data Science in Your Pocket's Medium analysis: a Single-Pass mHC for residual stream mixing, the Engram conditional memory module (196B sparsely accessed parameters), and DSpark speculative decoding with confidence-scheduled verification. The whole thing is natively multimodal, with built-in visual understanding.

My take: this is the first flagship architecture from the major labs openly optimized for the agent loop — massive re-reading, marginal generation — rather than for the chatbot. The rest of the market will follow, probably before the end of 2027.

For context on this release in the V4 lineage, our article DeepSeek V4: two new models — Pro and Flash — change the game traces the shift. The release itself is covered in detail in DeepSeek V4.1-Flash, a new open-weight model available today.

890 bytes per token: the frontal assault on the KV-cache

The KV-cache was the last physical bottleneck for long contexts. V4.1-Flash reduces it by a factor of 4 compared to V4-Flash, and by 437x compared to DeepSeek V1.

The main mechanism is called Compressed Sparse Attention 2 (CSA2), and its elegance lies in a simple decision: each layer is statically assigned one of three modes. Full mode: the layer computes its own KV. Reindex mode: it reuses the KV and the K indexer of the last Full layer, rescoring with its own Q indexer. Reuse mode: it also reuses the Top-K indices. According to MarkTechPost, there are 18 CSA2 layers at a compression ratio of 2, and 20 decoder layers at ratio 1, with a Hierarchical Sparse Indexer maintaining a candidate pool of up to 16,384 positions.

These are joined by an FP4 KV cache and SWA Bounded Replay, which reduces the persistent footprint on SSD or host memory to roughly 1/8 of the previous generation. The official DeepSeek docs confirm: 1/4 of the HBM, 1/8 of the storage.

The number that really matters, calculated by Agenteum: 890 bytes of KV per token in HBM. A 1M-token context therefore fits in ~0.93 GB of KV cache, versus ~2.15 GB naively in FP16. The max_position_embeddings value is confirmed at 1,048,576 in the config.json — this isn't marketing, it's measurable.

The technical paper is available on alphaXiv for anyone who wants the implementation details.

The real revolution is in the pricing: 98.6% of your bill is cache

Agents generate almost nothing. They re-read everything.

The most revealing measurement comes from Agenteum: over 30 days of traffic from a real agent, 26.7M fresh input tokens, 5.0M output — and 1.92 billion cached input tokens, i.e., 98.6% of everything sent. An agent that reworks the same repository, the same tool definitions, or the same conversation history re-reads its entire prefix on every turn.

Here is V4.1-Flash's pricing table at launch:

Off-peak is 50% of peak, with two peak windows on weekdays: 01:00-04:00 and 06:00-10:00 UTC, i.e., business hours in China. Concurrency limits increase from 500 to 2,500 requests.

Now, the comparison that stings. On 50M of tokens re-read as cache-read, VentureBeat calculates: ~$0.15 on V4.1-Flash off-peak, versus $15 on Kimi K3, $20 on GPT-5.6 Sol, and $25 on Claude Opus 5 at their published cache-read rates. A factor of 100 to 170 on the line that actually dominates your agent bills.

And the head-to-head: at identical traffic, the same agent workload costs $12.78 on DeepSeek off-peak versus $64.00 on GLM-5.3-Flash — a 5x gap that comes down to a single row in the table, the 10x-lower cache-input rate.

The methodological lesson, hammered home by VentureBeat: measure your cache-hit ratio and your cost per completed task, not the headline input price. The OpenDesign case study sums it all up: V4.1-Flash reaches 98% of GPT-6 Astra's quality score for 1.4% of its cost. If you're looking for the best model for your agents today, our comparison of LLMs for AI agents details these trade-offs.

DeepSeek vs GLM-5.3-Flash: two opposing answers to the same problem

Two labs attacked the same ailment — long context and unaffordable hidden input — from opposite directions. DeepSeek compressed the KV-cache (CED + CSA2 + MXFP4 + SWA Bounded Replay). Z.ai simplified the attention matrix: a hybrid linear + sparse MLA architecture with no positional encoding, which reduces attention compute by a factor of 3 and the KV cache by a factor of 4.4 compared to GLM-5.3.

The comparison table, according to Yotta Labs:

The batch scenarios from Intelligent Living are decisive: 10M input + 2M output uncached off-peak, that's $2.70 vs $2.50 — virtually a tie. At 80% cached input off-peak: $1.52 vs $1.54. But at peak: $5.40 vs $2.50 — GLM wins.

The verdict by traffic profile: cache-heavy and off-peak clearly favors DeepSeek; flat, predictable and daytime favors GLM. On the speed front, V4.1-Flash outputs ~194-198 tokens/s versus ~90 for GLM. On the hardware front, on an 8x H100 node, GLM leaves ~300 GB for KV and batch, V4.1-Flash ~130 GB — but the persistent cache at 890 bytes/token means a full 1M context fits under one gigabyte.

September 14: your endpoint swaps brains without warning

Starting at 12:00 PM Beijing time on September 14, 2026, all requests to deepseek-v4-pro — a retired model — will route to V4.1-Flash, at the lowest unit price, until V4.1 Pro is released. And that's not all: the old IDs deepseek-v4-flash and deepseek-v4-flash-vision-exp remain callable but now resolve to the new model, merging text and image under a single API entry point. The details of the switch are in our article DeepSeek swaps V4-Pro for V4.1-Flash behind the same endpoint: why you should re-test your pipelines before September 14.

Why this matters: Agenteum documents the case of an agent losing 30% of useful context due to context tables hard-coded with the old IDs. The model behind the endpoint changed, but your code still assumes yesterday's behavior.

Three checks before the 14th: remove hard-coded model IDs from your context tables, re-verify your context window assumptions (1M confirmed, 384K max output, with JSON output, tool calling, and the Responses API on the menu), and re-run your agentic test suites against the new model. "Thinking" is enabled by default with three effort levels (low/high/max) — a silent behavior change that can catch your output parsers off guard.

My architect's advice: slip a thin proxy between your code and the DeepSeek API, with the model mapping in a config file. A simple VPS from Hostinger is enough for this relay, and the next endpoint switch will become a one-line change instead of an emergency operation.

The silent migration remains a debatable choice in terms of communication. But no one is going to complain, because third-party testing already ranks V4.1-Flash ahead of V4-Pro on performance, cost, speed, and total runtime. Flash kills Pro, literally.

Benchmarks: Flash beats the flagship it replaces

Yes, a model branded "Flash" outperforms the Pro model it replaces. That's the most notable reversal of this release.

The sourced numbers: a Codeforces rating of 3471 (MarkTechPost), results ahead of Claude Opus 5 on Terminal Bench 2.1 and DeepSWE (Medium), and a DeepSWE v1.1 of 74.2 versus 63.4 for GLM-5.3-Flash (Yotta Labs).

On the agentic benchmarks from vals.ai, the gap widens: +54 points on Vibe Code Bench, +33 on ProofBench, +25 on Code Migration versus GLM. Agentic execution is this model's natural playground.

Let's be honest about the limitations: GLM remains the better generalist (Finance Agent, Legal, MedScribe). And GLM's "announced 1M context" has only been evaluated up to 300K — so keep a general skepticism toward advertised context windows until you've tested them on your own workloads.

To situate V4.1-Flash in the current landscape, our comparison of the best LLMs for coding is updated every month.

What about running it locally? 510 GB won't fit on your Mac

The weights are MIT and open, but serve them with a cluster, not a laptop. The FP8 checkpoint weighs in at ~510 GB across 48 shards; even on an 8x H100 node, only ~130 GB is left for the KV cache and batch after loading the weights. The asymmetric CED architecture also requires dedicated engine support — your usual runtime won't handle it out of the box.

The precedent is encouraging, however: after the release of V4 Flash, antirez launched ds4, a local inference engine making DeepSeek V4 Flash usable on a Mac. Expect quantized community builds for V4.1-Flash in the weeks ahead, but check that CED support is actually implemented before porting your agents to them.

In the meantime, if you want local right now, our guide to setting up a local LLM covers Ollama and LM Studio, and the comparison of the best local LLMs lists the models that fit on consumer hardware.

For 95% of agent use cases, the honest answer is: go with the API. At $0.003 per million cached tokens, the self-hosting math doesn't even enter the picture.

❌ Common Mistakes

Mistake 1: Comparing displayed input prices

The headline price ($0.15/1M) is not your real price. On an agent, 98.6% of tokens come from the cache. Measure your cache-hit ratio over 30 days and calculate a cost per completed task — that's the only metric that predicts your bill.

Mistake 2: Leaving model IDs hardcoded

Documented real-world case: 30% of usable context lost after an endpoint switch, because context tables were frozen on the old IDs. Centralize model mapping in a config, and retest your pipelines before September 14.

Mistake 3: Running heavy batches right in the middle of peak hours

Peak windows are 01:00–04:00 and 06:00–10:00 UTC on weekdays. A batch of 10M input + 2M output costs $2.70 off-peak versus $5.40 at peak. Schedule your heavy jobs during off-peak hours — the rate there is half as much.

Mistake 4: Trusting the advertised context window without testing

GLM-5.3-Flash advertises 1M but is rated at 300K. V4.1-Flash confirms 1,048,576 in its config.json — but "supported" doesn't mean "fully performant." Test your own long workloads before committing to production.

❓ Frequently Asked Questions

Will my code break on September 14?

No, not outright: requests to deepseek-v4-pro automatically route to V4.1-Flash at the lowest unit price, until V4.1 Pro is released. But if you have hardcoded IDs, context tables, or parses tied to the old model, retest — one agent lost 30% of usable context for exactly this reason.

Can I use V4.1-Flash commercially?

Yes. The weights are MIT-licensed and published on Hugging Face on launch day. Distillation, fine-tuning, commercial product: everything is allowed. Expect ~510 GB FP8 and serious multi-GPU hardware to serve the model yourself; otherwise, the API or Baseten get the job done.

Why is my agent bill dominated by cache?

Because an agent resends its entire prefix every turn: code repo, tool definitions, conversation history. Over the 30 days measured, 1.92 billion cached tokens vs 26.7 million fresh. Hence the decisive importance of the $0.003/1M off-peak cache-hit rate.

V4.1-Flash or GLM-5.3-Flash?

Input-heavy traffic, shiftable to off-peak: DeepSeek, up to 5x cheaper on real workloads. Flat, predictable, mostly daytime traffic: GLM, with its flat output pricing at $0.50/1M. GLM also remains the better generalist on vals.ai (Finance Agent, Legal, MedScribe).

When is V4.1 Pro coming out?

No date announced. Until then, the deepseek-v4-pro endpoint serves V4.1-Flash at the lowest price. Tests from several parties cited by DeepSeek already place Flash ahead of Pro on performance, cost, and speed — so the urgency for V4.1 Pro isn't obvious.

✅ Conclusion

DeepSeek V4.1-Flash demonstrates that the next LLM war will be fought over the KV-cache byte and the "cache-hit" line in the pricing, not over benchmark tables. Before September 14, retest your pipelines — then compare the results on our monthly comparison of the best LLMs.