DeepSeek's 'minor' Flash update adds eleven points on Terminal-Bench
DeepSeek calls its October 17 Flash checkpoint a minor update. Its own table lifts Terminal-Bench 4.0 from 31.2 to 42.3 and still trails Opus 5 on HLE.
Vlad Makarovreviewed and publishedDeepSeek pushed DeepSeek-V4.1-Flash-1017 to Hugging Face in the early hours of October 17 with no launch post and no blog post: forty-eight safetensors shards, an MIT licence and a model card. The card's opening sentence calls it "a minor update to DeepSeek-V4.1-Flash," the open-weight model we covered in September. Its own table then moves Terminal-Bench 4.0 from 31.2 to 42.3 — eleven points, three weeks later, on a release the vendor has labelled minor.
Both sentences are DeepSeek's. Only one can be checked, and not on release day.
Three weeks is a short time for eleven points
The card does not stop at "minor." The same paragraph claims "substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks," says the model cuts task execution time and output length, and argues it advances "the performance-cost Pareto frontier across multiple agent harnesses." A release that is minor in versioning and substantial in measured capability is strange to publish without a word about what changed: no training description, no data note, and the linked technical report belongs to the September model.
That absence matters because the size of the jump is the story. Terminal-Bench 4.0 rises 11.1 points, ProgramBench 8.8, Apex Bench 7.5 and Automation-Bench 5.7, while DeepSWE v1.1 — where September was already strong at 74.2 — gains 2.8. That pattern fits targeted work on agent scaffolding better than a bigger pretraining run, but the card never says so, and it does not state whether the September column used the same harness versions. It notes only that Terminal-Bench ran under Claude Code and DeepSWE under Mini-SWE-V2.
The table, and which way each comparison runs
| Benchmark | 1017 | V4.1-Flash | K3 | GLM-5.3 | GPT-5.6 Sol | Opus 5 |
|---|---|---|---|---|---|---|
| HLE w/o tools | 42.6 | 36.8 | 43.5 | 42.0 | 44.5 | 56.3 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | 12.6 | 37.9 | 39.9 | 51.8 |
| DeepSWE v1.1 | 77.0 | 74.2 | 67.5 | 66.9 | 73.0 | 74.0 |
| ProgramBench | 29.1 | 20.3 | 17.5 | 19.0 | 23.0 | 37.0 |
| NL2Repo-Bench | 68.2 | 65.4 | 58.0 | 58.0 | 56.8 | 75.3 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | 46.7 | 48.8 | 45.8 | 50.3 |
| Agents' Last Exam | 35.4 | 31.8 | 27.6 | 28.5 | 26.7 | 28.6 |
| Apex Bench v1.1 | 72.1 | 64.6 | 47.8 | 56.6 | 54.3 | 65.8 |
Against Opus 5, DeepSeek wins four rows outright — DeepSWE by 3.0, Automation-Bench by 10.2, Agents' Last Exam by 6.8, Apex Bench by 6.3 — and loses the ones closest to general reasoning: HLE by 13.7, Terminal-Bench by 9.5, NL2Repo-Bench by 7.1 and ProgramBench by 7.9. GPT-5.6 Sol leads only on HLE, by 1.9, and trails on the other seven. The vision rows, cut from the table, improve on September and fall short of Opus 5: Chartography 83.0 against 78.9, BabyVision 92.5 against 89.6, ZeroBench-main 51.0 against 49.0, with GLM-5.3 absent from all three.
K3's 12.6 on Terminal-Bench, beside a 43.5 on HLE, is either a very weak harness result or a measurement the card does not describe.
The parameter count in the weights is not the one in the prose
Hugging Face's sidebar prints "763B params" for the repository, and the API agrees: the safetensors totals sum to 763,205,315,794, split across I8 at 557.2B, F8_E4M3 at 204.0B, BF16 at 2.0B and F32 at 42.3M, with 510 GB of used storage across 48 shards. DeepSeek's card and the OpenRouter listing describe something else: a 552B-parameter backbone activating 8B parameters on input and 16B on output.
Both numbers are real and they are not the same quantity. A sparse mixture of experts stores every expert and spends a fraction per token, which explains 552B total against 16B active, not 552B against 763B. The mixed-precision tensor total counts what is on disk, and a checkpoint storing some components in more than one form — quantization scales, a duplicated output head, a speculative-decoding module — will sum higher than the architecture. DeepSeek has published no tensor-by-tensor accounting: 763B is a storage figure, 552B the architecture figure. Neither settles the other.
The price is the actual release
OpenRouter's cheapest standard route is Morph at $0.09 per million input tokens and $0.55 per million output, cache reads at $0.008. DeepSeek's own endpoint lists $0.15 and $0.60 with cache reads at $0.003, at 1.0M context and roughly 155 tokens per second. The page also publishes what customers actually pay: a weighted average input price of $0.0178 against a $0.15 list, on a 96.1% cache-hit rate at DeepSeek's endpoint.
That gap is the product. The card's cited mechanisms — KV caching at roughly a quarter of the previous generation's memory, and shorter outputs — are cost levers, and it puts them ahead of the scores: "enabling users to complete tasks faster and at lower cost." At these numbers an agent looping through a hundred thousand tokens pays cents, which explains the traffic OpenRouter reports better than any score does.
No chat template, and that is the interesting part
The release ships without a Jinja chat template. In its place is an encoding/ folder holding a self-contained Python reference implementation with tests for multi-turn conversations, tool calling, thinking mode, numeric reasoning effort, mid-conversation system messages and interleaved image content, plus deepseek-recipe, Rust libraries with Python bindings that encode Messages, Chat Completions and Responses requests into the same prompt format.
Recommended sampling is temperature 1.0, top_p 1.0, a 1M-token context and at least 256K output tokens. A card that ships a reference encoder and asks for 256K tokens of headroom is saying the model is built for long agent runs, and that reproducing its scores means matching the prompt format exactly. Every harness wrapping it has to port code rather than load a template file.
What would settle it
Nothing in that table has an independent counterpart yet. Artificial Analysis carries a V4.1 Flash entry — 39 on its Intelligence Index, 552B parameters, 16B active — but the page is dated to the September release and its weights link points at the older repository, not this checkpoint. One number still matters: OpenRouter's benchmark block, sourced from Artificial Analysis, records 26.8% on Terminal-Bench 4.0 for the September model, against the 31.2 DeepSeek credits that same model with. A four-point gap between a vendor column and an outside harness is not a scandal, but it is the right size of skepticism to bring to a claimed eleven-point jump.
What the repo establishes is less exciting and more durable. The weights are downloadable, ungated, MIT-licensed and 510 GB, with more than 880 likes on the page by the following morning — attention, not evidence. The licence is the materially open part of this release, and it has not changed. Whether 42.3 is a real number will be settled by someone running Terminal-Bench against the checkpoint without DeepSeek in the loop.
