DeepSeek's official Hugging Face model card for V4.1 Flash is a different document from the launch thread this blog covered on September 10: where the thread led with a benchmark chart against named rivals, the model card leads with architecture, then backs it with a base-model table, a frontier-comparison table, and — unusually for a launch document — a table showing how much the model's own scores move depending on which agent harness runs it. That last table is the most interesting thing in the release, and it's worth reading closely rather than skimming past.
A decoder that doesn't compute its own attention cache
The headline architectural claim is a Causal Encoder-Decoder (CED) design: a 40-layer transformer split into a 20-layer causal encoder and a 20-layer decoder, where the decoder's global key-value cache is projected from the encoder's final hidden states rather than derived fresh from each decoder layer. That's the mechanism behind the model's most striking efficiency number: only 8B of the 552B backbone activate per token during prefill, versus 16B during decode.
That asymmetry is a deliberate bet on what agentic workloads actually look like. Prefill is the phase that processes the prompt — the tool outputs, file contents, and conversation history an agent reads before it writes anything — and in a long-running agent loop, reading dwarfs writing. By making prefill the cheap phase (8B active) and reserving more capacity for decode (16B active, where output quality matters most), the architecture targets exactly the workload distribution this blog's own coverage of DeepSeek's agentic benchmarks has tracked all year: agents spend most of their tokens reading, not generating.
The rest of the cache-efficiency story is Compressed Sparse Attention 2 (CSA2), which assigns each attention layer one of three fixed roles — Full, Reindex, or Reuse — so that most layers share key-value state and sparse-attention indices computed by a small number of Full layers rather than recomputing their own. A Hierarchical Sparse Indexer restricts later layers to a candidate pool built by the first Full layer, which is what keeps indexing cost from growing with context length even out to the model's 1M-token window. Paired with FP4 key-value caching (an E2M1 format with one E4M3 scale per 16 channels), this gets the model's global KV cache footprint to 890 bytes per token — a 4-fold reduction from DeepSeek-V4-Flash and a 437-fold reduction from the original DeepSeek-V1, by the card's own figures. A separate technique, SWA Bounded Replay, attacks a different cache cost: it reconstructs sliding-window attention state by replaying only the most recent tokens instead of persisting it to disk, cutting the persistent KV footprint to roughly an eighth of V4-Flash's. These are two distinct numbers about two distinct caches — the 4x/437x figures describe the global KV cache, the roughly 8x figure describes what has to be kept in persistent storage — and the card is careful to keep them separate even though both compound into the same story: a much smaller memory footprint per token than any prior DeepSeek generation.
Two more components round out the architecture, both stated plainly and neither independently verified elsewhere: an "Engram" conditional memory bank of 196B parameters, accessed sparsely through token-based lookup rather than dense computation, and DSpark, a speculative-decoding scheme using semi-autoregressive draft generation with confidence-scheduled verification. The card doesn't say how many of Engram's 196B parameters get touched per token, only that access is sparse — worth noting as an open question rather than filling in a number that isn't there. Taken at face value, Engram sits outside the 552B backbone figure entirely, meaning the model's total addressable parameter count is closer to 748B even though the widely-quoted number will be the smaller backbone figure.
One more component connects to a story this blog has already told. Z.ai's GLM-5.3-Flash reveal last month included an independent teardown by ML educator Sebastian Raschka, who attributed GLM's residual-stream mechanism to "Manifold-Constrained Hyper-Connections" (mHC) in "a DeepSeek V4-style four-stream design" — his own reconstruction, not a confirmed spec at the time. DeepSeek's own card for V4.1 Flash now names the same abbreviation directly: "Single-Pass mHC," described as revised residual-stream mixing run through what the card calls an "efficient Mega-mHC kernel." That's DeepSeek confirming, in its own document, that mHC is indeed a DeepSeek-originated technique — Raschka's attribution holds up, and V4.1 Flash is DeepSeek's own next iteration on it, not a response to anyone borrowing it.
The base model already shows the pattern the instruct model would repeat
Before any post-training, DeepSeek-V4.1-Flash-Base already displays the tradeoff that would carry through to the final model. Against DeepSeek-V4-Pro-Base — a much larger 1.6T-parameter backbone activating 49B params per token, more than six times V4.1-Flash-Base's activation — the smaller model wins or ties on code and math (HumanEval 79.4 vs. 76.8, BigCodeBench 60.6 vs. 59.2, GSM8K 93.0 vs. 92.6) while trailing clearly on knowledge recall and long context (SimpleQA-Verified 42.3 vs. 55.2, MultiLoKo 45.5 vs. 50.9, LongBench-V2 45.2 vs. 51.5). That's the same split that shows up in the instruct-model comparison below, which means it isn't an artifact of post-training choices — it's already present in the pretrained backbone, before any reinforcement learning or distillation touches it.
What the frontier-comparison table actually shows
DeepSeek's comparison against Opus 5, GPT-5.6 Sol, Kimi K3, GLM-5.3, and its own V4-Pro and V4-Flash confirms the numbers already reported from the September 10 launch thread: DeepSWE v1.1 at 74.2 against Opus 5's 74.0, CyberGym at 88.1 ahead of every named rival, Terminal-Bench 3.0 clearly behind Opus 5 (30.0 vs. 43.3), and GPQA Diamond and HLE both led by DeepSeek's own V4-Pro (92.4 and 42.7, against V4.1 Flash's 90.9 and 36.8). One result the launch thread didn't foreground: on AutomationBench, V4.1 Flash's 54.8 beats Opus 5's own 50.3, not just its cheaper open-weight rivals — a genuine lead over the frontier model DeepSeek was benchmarking against, not just a near-tie.
The three vision-flavored benchmarks — Chartography, BabyVision, and ZeroBench-main, all run "with tools" — confirm the number this blog cited from Sakana's Fugu launch three days later: Opus 5 at 84.0 on Chartography is exactly the figure this card reports, in the same "Chartography w/ tools" row. Worth being precise about what this table does and doesn't include: it has no row for Claude Fable 5.1 at all — only Opus 5 appears as Anthropic's representative — so any Fable 5.1 number circulating elsewhere for this benchmark did not come from DeepSeek's own chart.
The table that complicates the headline
Most model cards stop at the comparison table above. This one adds a second table, evaluating the same model on the same two benchmarks — DeepSWE v1.1 and Terminal-Bench 2.1 — across eight different agent scaffolds: Claude Code, Codex, OpenCode, Pi, mini-SWE, and three modes of DeepSeek's own harness (Minimal, Standard, PTC).
On DeepSWE v1.1, the spread runs from 65.5 (OpenCode) to 74.2 (mini-SWE) — 8.7 points of variance on the identical model and identical benchmark, produced entirely by which scaffold executes the agent loop. On Terminal-Bench 2.1, the range is narrower but still real: 84.1 (Codex) to 90.6 (DeepSeek Harness Minimal).
Here's the detail worth sitting with: 74.2, the DeepSWE v1.1 score used in the frontier-comparison table to say V4.1 Flash "edges past" Opus 5's 74.0, is the mini-SWE result — and mini-SWE is also the single highest-scoring scaffold of the eight tested. The methodology note explains the choice as aligning with DeepSWE v1.1's own official setup requirements, which is a legitimate reason to pick a specific harness rather than an arbitrary one. But it means the reported margin over Opus 5 — two-tenths of a point — is more than seven times smaller than the gap between V4.1 Flash's own best and worst scaffold on the same task. Run the model through DeepSeek's own Minimal harness instead of mini-SWE, and the score drops to 72.6 — a clear loss to Opus 5's 74.0, not a win. On Terminal-Bench 2.1, the same pattern holds in miniature: 90.6, the number in the frontier table, is DSH Minimal's result, and DSH Minimal is (by a hair, over mini-SWE's 90.3) the best of the eight scaffolds tested there too.
None of this means the headline number is fabricated or the methodology note is dishonest — DeepSWE v1.1 genuinely may have an official recommended scaffold, and DeepSeek disclosed the full eight-way comparison in the same document rather than hiding it, which is more transparency than most launch benchmarks offer and exactly the kind of table that lets a reader check the claim instead of taking it on faith. But the plain reading is this: the single number chosen to support "edges past Opus 5" is also, whether by methodology or by coincidence, the ceiling of what that model scores across every scaffold DeepSeek itself tested. A margin smaller than a model's own harness-to-harness noise floor is not a meaningful lead, and this is the rare model card that hands you the data to see that for yourself.
Practical notes for anyone actually running it
The release ships without a Jinja chat template, instead providing a self-contained Python reference (encoding.py) covering multi-turn conversations, tool calling, thinking mode, numeric reasoning effort, mid-conversation system messages, and interleaved images. For production use, DeepSeek separately released deepseek-recipe, Rust libraries with Python bindings that convert Messages, Chat Completions, and Responses API requests into prompts or token IDs and parse the model's output back out — leaving inference, tool execution, and HTTP transport to the caller. The model exposes a continuously controllable reasoning-effort setting from 1 to 100 rather than a small number of discrete modes, trading inference cost for accuracy on a single dial; all benchmark numbers in the frontier-comparison table use the maximum setting, 100.
What to expect next
- Watch for independent reproductions of the DeepSWE v1.1 gap. DeepSeek's own eight-scaffold table already shows the margin over Opus 5 is scaffold-dependent; an outside evaluator running both models under one shared, neutral harness would settle whether "edges past Opus 5" survives outside DeepSeek's own preferred setup.
- Watch whether other labs start publishing scaffold-sensitivity tables of their own. This one is unusually candid about a source of variance most launch benchmarks never disclose; if it becomes a norm rather than an outlier, that's a genuine improvement in how agentic benchmarks get reported industry-wide.
- Watch for third-party confirmation of the Engram memory component. A 196B-parameter sparse lookup structure sitting alongside a 552B backbone is a large, unusual claim with no stated per-token activation count; independent interpretability or efficiency analysis would be the next real check on it.
- Watch whether "Single-Pass mHC" gets its own teardown, the way GLM-5.3-Flash's architecture did from Sebastian Raschka. DeepSeek's card confirms the technique's origin but doesn't fully specify the "Mega-mHC kernel" beyond naming it, leaving room for the same kind of independent reconstruction that clarified GLM's version.