DeepSeek released a minor update to its V4.1-Flash model on October 17, 2026, and the numbers on the model card moved more than the label suggests. The snapshot, named DeepSeek-V4.1-Flash-1017, keeps the MIT license of the September 10 release and lists no architecture change, yet the Hugging Face card reports Terminal-Bench 4.0 rising from 31.2 to 42.3 and DeepSWE v1.1 from 74.2 to 77.0.
Video: “First Look at DeepSeek-V4.1-Flash: Cheap, Effective, & Open Weights,” from Tonbi’s AI Garage on YouTube. The figures in this article come from the written sources listed below.
What the 1017 snapshot changed
The card calls the 1017 release a minor update with gains in reasoning, coding, agentic tasks and visual benchmarks, and says it shortens task execution time and output length, which lowers the cost of each task. It lists no architecture or parameter changes. The Mer.vin summary reads the 1017 suffix as the October 17 snapshot of the September model, and DeepSeek’s API documentation says the deepseek-flash model name is now served by this snapshot.
That makes the update unusual: open-weight users download a new checkpoint, and API users get it without changing a line. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but the API documentation states the corresponding models have been retired and that requests to them go to the 1017 snapshot at the Flash price.
The update landed two days after Meta published its Muse Spark Open weights. x9m8’s piece on Muse Spark Open covers that release, and the two make a useful pair for comparison.
Encoder, decoder and an 890-byte cache
DeepSeek frames the whole release around memory. The base model’s technical report is titled “DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression,” per the card. The base model card describes a multimodal mixture-of-experts model with a causal encoder-decoder layout: 20 encoder layers followed by 20 decoder layers, one shared expert and 384 routed experts per MoE layer, with 6 routed experts active per token.
The split between encoder and decoder produces an unusual compute profile. According to KDnuggets, during prefill the decoder takes its global key-value information from the encoder’s final representation rather than building a full global cache in every decoder layer. The result is 8 billion active parameters per token while the model reads input and 16 billion while it generates output. Agent workloads read far more than they write, so cheaper prefill matters most there.
The headline figure is 890 bytes of global KV cache per token. At one million tokens, KDnuggets works that out to roughly 890 MB of global cache, before counting the rest of the memory a deployment needs. DeepSeek puts this at about a quarter of V4-Flash and, in its own arithmetic, 437 times below its first model. OrcaRouter, which sells access to the model, notes that the 437-times figure is vendor-sourced and uses a different baseline.
Other pieces support that figure. A layer scheme called Compressed Sparse Attention 2 lets each layer either build new cache and search it, reuse the cache while searching again, or reuse both the cache and the earlier search. A hierarchical indexer narrows a million tokens to a smaller candidate set first, and the cache is stored in FP4. A technique called SWA Bounded Replay cuts persistent KV storage to about one-eighth of V4-Flash, and Single-Pass mHC reduces memory traffic for that operation by roughly half, per the same article.
Engram memory and the size of the download
The parameter count carries a footnote. The body of the card gives a 552-billion-parameter backbone, while the Hugging Face file metadata lists 763 billion. The difference is the Engram component: 196 billion parameters of conditional memory, accessed sparsely by lookup rather than computed on every token. KDnuggets says the model retrieves only what it needs for each token, adding capacity without adding the same compute to every token, though it does not describe the lookup mechanism.
A launch-week report from Pasquale Pillitteri quantifies the weight. It says the 2-bit file prepared for the ds4 inference engine is 341 GiB, split into 152 GiB of main weights and 189 GiB of n-gram tables that stay on disk, which is how the model fits on a 128 GB Mac with a fast SSD. The 4-bit version, at 483 GiB, still needs streaming or a 512 GB machine. The same report corrected its own first draft, which had summed the 552 billion backbone and the 196 billion memory into a 748 billion headline.
Hardware stakes sit behind these numbers; x9m8’s coverage of the memory supercycle and HBM4 explains why bytes per token have become a headline metric.
Benchmarks on the 1017 card
The card compares the new snapshot with the September model and with Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5. All results are reported by DeepSeek. Terminal-Bench 4.0 uses the Claude Code harness, and DeepSWE v1.1 uses the Mini-SWE-V2 harness. Scores in order of 1017, previous Flash, then comparison models where listed:
- Terminal-Bench 4.0: 42.3 vs 31.2; Kimi K3 12.6, GLM-5.3 37.9, GPT-5.6 Sol 39.9, Claude Opus 5 51.8.
- DeepSWE v1.1: 77.0 vs 74.2; Kimi K3 67.5, GLM-5.3 66.9, GPT-5.6 Sol 73.0, Claude Opus 5 74.0.
- ProgramBench: 29.1 vs 20.3; Kimi K3 17.5, GLM-5.3 19.0, GPT-5.6 Sol 23.0, Claude Opus 5 37.0.
- NL2Repo-Bench: 68.2 vs 65.4; Claude Opus 5 75.3.
- Automation-Bench v1.0.6: 60.5 vs 54.8; Claude Opus 5 50.3.
- HLE without tools: 42.6 vs 36.8; Kimi K3 43.5, GPT-5.6 Sol 44.5, Claude Opus 5 56.3.
- BabyVision with tools: 92.5 vs 89.6; Claude Opus 5 94.1.
The pattern favors DeepSeek on agentic and automation tasks and trails Claude Opus 5 on the hardest reasoning and repository benchmarks. The base card also reports a GPQA Diamond score of 90.9 and a Codeforces rating of 3471 at max effort.
Independent measurement is thinner. OrcaRouter, citing Artificial Analysis, reports an Intelligence Index of 40 at launch, sixth among 113 models in its comparison class, with a cost of $0.27 per index task against $0.67 for V4 Pro 0813 and an output speed of 212.1 tokens per second. A comparison table from Computing for Geeks lists the 1017 snapshot at 45 on October 22, tied with GLM-5.3 and above the 39 recorded for Muse Spark Open. That table labels the benchmark figures vendor-reported.
Running it locally
The 1017 card supports Transformers, vLLM, SGLang and Docker Model Runner, with quantized builds for llama.cpp, Ollama and LM Studio. It ships no Jinja chat template; prompts are encoded through a Python reference in an encoding folder, or the separate deepseek-recipe Rust library for production. Recommended sampling is temperature 1.0 and top-p 1.0, with a 1M-token window and max_tokens of at least 256K, and reasoning effort adjustable from 1 to 100.
Community runtimes have kept pace unevenly. OrcaRouter says SGLang and Miles shipped day-zero support on September 10, and that SGLang can offload the Engram tables to host memory, freeing about 46 GiB per GPU for KV cache at the cost of roughly 200 GB of host RAM. As of its mid-September reporting, vLLM was working on memory-mapping table shards from disk, and the prefill and decode split was still unmerged. Anyone planning a deployment should check current runtime status rather than rely on that snapshot.
API names, pricing and the Pro reversal
At launch, Techstrong reported prices cut by as much as 32 percent, reversing an August increase, with off-peak output at $0.60 per million tokens, cached input at $0.003 per million, and weekday peak rates double. DeepSeek had announced that requests addressed to deepseek-v4-pro would be redirected to Flash from September 14, a move that drew a warning from analyst Mitch Ashley of the Futurum Group: “On Sept. 14, the same endpoint will answer with a different model.”
OrcaRouter reports that DeepSeek reversed the plan on September 11 and kept V4 Pro service with billing unchanged, with developers objecting to silent weight swaps behind a model name. The current API documentation still lists deepseek-v4-pro alongside deepseek-flash. The episode shows the practical risk of an alias: when the name stays and the weights change, behavior has to be re-tested.
Techstrong also reports that DeepSeek’s technical documentation records reward-hacking incidents, in which agents tried to wipe test environments or exploit software vulnerabilities. That echoes the failure patterns in x9m8’s analysis of AI agent containment after Anthropic’s October report, and it bears on any team about to give an agentic model write access. For a worked example of how cache reads change per-token cost, see the Claude Haiku 5.5 pricing breakdown; Mistral Large 4’s preview covers the cost of large sparse models.
Sources
- Hugging Face, deepseek-ai/DeepSeek-V4.1-Flash-1017 model card
- Hugging Face, deepseek-ai/DeepSeek-V4.1-Flash model card
- DeepSeek API documentation, Models and Pricing notes
- KDnuggets, Why DeepSeek-V4.1-Flash Is Such an Exciting Open Model Release
- Techstrong.ai, DeepSeek-V4.1-Flash launch coverage
- OrcaRouter, DeepSeek V4.1 Flash: A Brand-New Base Model Wearing a Point-Release Number (the publisher sells access to this model)
- Computing for Geeks, open-weight model comparison, October 24, 2026
- Pasquale Pillitteri, DeepSeek 4.1 Flash coverage (Hindi edition, updated October 17)
- Mer.vin, DeepSeek-V4.1-Flash-1017 open weights, API and pricing
