DeepSeek V4.1-Flash-1017: +11.1 on Terminal-Bench, MIT Weights, Same Token Price

TL;DR

DeepSeek shipped V4.1-Flash-1017 on October 17, a checkpoint update to the September V4.1-Flash. Same API identifier (deepseek-flash), same token price, MIT weights on Hugging Face and ModelScope.
Terminal-Bench 4.0 goes from 31.2 to 42.3 (+11.1 points, +35.6% relative). ProgramBench from 20.3 to 29.1 (+43.3% relative). Eleven published evaluations, all up.
Against GPT-5.6 Sol, DeepSeek wins seven of eight published comparisons — but Claude Opus 5 still leads Terminal-Bench (51.8) and ProgramBench (37.0).
The interesting number is not the gain, it is the cost per success: DeepSeek reports shorter outputs and faster completion at unchanged prices. For agent pipelines, that is the metric that decides.

A checkpoint, not a new model

There is no new architecture here. V4.1-Flash-1017 keeps the asymmetric Causal Encoder–Decoder MoE design introduced in September: a 552B-parameter backbone with roughly 8B parameters active during input and 16B during output, a 1M-token API context, 384K maximum output, text and image input, thinking mode and tool calls. (Source : Data Studios — DeepSeek V4.1 Flash 1017: Open Weights, Benchmarks and API Pricing)

DeepSeek announced it on X as a “minor update” with “substantial improvements across reasoning, coding, agentic tasks and visual benchmarks.” Minor is doing a lot of work in that sentence, because the deltas are not minor on the benchmarks that matter for agents. (Source : AI Understanding — DeepSeek releases V4.1-Flash-1017)

The migration cost is zero: applications pointing at deepseek-flash get the new checkpoint without touching a line of configuration. That is the underrated fact in the release. A model upgrade that requires a provider swap, a new SDK and a re-run of every eval is a project. One that arrives silently behind the same API string is a Tuesday.

The numbers, and what they actually say

DeepSeek published eleven evaluations, all improved over the September checkpoint:

HLE, no tools: 36.8 → 42.6
Terminal-Bench 4.0: 31.2 → 42.3 (+11.1)
DeepSWE v1.1: 74.2 → 77.0
ProgramBench: 20.3 → 29.1 (+8.8)
NL2Repo-Bench: 65.4 → 68.2
Automation-Bench v1.0.6: 54.8 → 60.5
Agents’ Last Exam: 31.8 → 35.4
Apex Bench v1.1: 64.6 → 72.1
Chartography, with tools: 78.9 → 83.0
BabyVision, with tools: 89.6 → 92.5
ZeroBench-main, with tools: 49.0 → 51.0

The distribution is the story. Vision barely moves (+2 to +4 points), which suggests the multimodal stack is already near its ceiling. The gains concentrate exactly where an agent spends its money: terminal execution, program construction, structured multi-step automation. An 11.1-point jump on Terminal-Bench is the difference between an agent that reads a stack trace and gives up, and one that reads it, edits the test, and retries on a different assumption.

Against competitors, DeepSeek’s own table is flattering but not dishonest. On the eight benchmarks shown side by side, V4.1-Flash-1017 beats GPT-5.6 Sol on seven: Terminal-Bench 42.3 vs 39.9, DeepSWE 77.0 vs 73.0, ProgramBench 29.1 vs 23.0, NL2Repo 68.2 vs 56.8, Automation-Bench 60.5 vs 45.8, Agents’ Last Exam 35.4 vs 26.7, Apex 72.1 vs 54.3. It loses HLE without tools, 42.6 vs 44.5.

Claude Opus 5 is the harder wall: Terminal-Bench 51.8, ProgramBench 37.0, NL2Repo 75.3, HLE 56.3. The honest reading is not that open weights caught the frontier. It is that open weights closed the gap on execution-heavy agent work while staying roughly ten points behind on frontier reasoning.

One caveat matters more than the leaderboard: DeepSeek ran Terminal-Bench through the Claude Code harness and DeepSWE through Mini-SWE-V2. Tool access, retry policy and context management belong to the harness, not the model. Any comparison that ignores which scaffold ran which weights is measuring the scaffold too.

Why a checkpoint can buy eight points

DeepSeek credits reinforcement learning with self-feedback, particularly for tasks where the instruction is broad and the deliverable is a whole project. The published demos are a Portal 2-inspired level built in Godot, a 3D model of Beijing’s Forbidden City, small web games, a toy game engine.

Take the demos as demos. The mechanism is what survives scrutiny. Project-scale tasks fail cumulatively: an early architectural choice poisons every later step, and the agent has no external signal telling it which step was wrong. Self-feedback training teaches a model to notice that its own intermediate state contradicts the goal and to revise before continuing. That is a different skill from answering well, and it is the one benchmark suites have struggled to measure.

The transfer risk is real. A model trained to complete projects can also learn to declare a project complete. Every “built X from scratch” demo should be read next to a regression rate: how often does the same agent, on the same repo, quietly break something that used to work?

The cost-per-success question

Price per token is unchanged, which sounds like a non-event until you pair it with the other claim: shorter outputs, faster completion. If a task that consumed 40K output tokens now consumes 28K and finishes two turns earlier, the effective price of a completed task just fell by roughly a third without a single line on the pricing page changing.

That is the metric worth instrumenting before adopting the checkpoint. Not the leaderboard, not the demo reel — tokens per success and wall-clock per success, on your own tasks with your own tools. A benchmark gain that costs more retries is a regression dressed as progress; a flat benchmark with 30% shorter trajectories is a raise.

The counterweight is hosting. MIT weights mean you can run it yourself, but a 552B MoE backbone with 1M context is not a laptop download. The licence buys optionality, not independence: for most teams the practical choice stays the hosted API.

What to do with it

If you run agent workloads on deepseek-flash, the checkpoint already landed. Re-run your own evals before trusting the vendor table, and log three numbers per workload: success rate, output tokens per success, turns per success. Compare against your own last month of traces, not against Terminal-Bench.

If you are choosing a model for terminal-heavy automation, the ordering is clearer than it was: closed frontier models still lead hard reasoning, open weights lead price per outcome at roughly 80% of the ceiling, and the harness you wrap around either one moves the score by more points than a generation of weights.

FAQ

Is this a new model? No. It is a checkpoint update to V4.1-Flash, the September release. Same family, same API ID, same price, better weights.

Do I have to change anything to get it? If you call deepseek-flash, no — the new checkpoint is already served. If you self-host, pull the new weights under MIT.

Are the numbers independent? No. All results are developer-reported, and two benchmarks ran through third-party harnesses (Claude Code, Mini-SWE-V2), which makes them harness-sensitive.

Does it beat Claude Opus 5? On some agentic benchmarks, yes: Automation-Bench and Agents’ Last Exam. On Terminal-Bench, ProgramBench, NL2Repo and HLE, no. Opus 5 keeps a five-to-fourteen-point lead on the hardest ones.

What about the 1M context? Unchanged: 1M tokens in, 384K out, text and images. The checkpoint improves behaviour at that context, not its size.