DeepSeek V4.1 Flash-1017: What Changed, What It Costs, Should You Switch?
Answer first
DeepSeek V4.1 Flash-1017 is a dated snapshot update to DeepSeek's budget Flash line: MIT-licensed open weights, a 1M-token context window, and vendor-reported agentic-coding gains (Terminal-Bench 4.0 up from 31.2 to 42.3). API access starts at $0.05–$0.15 per million input tokens. It is worth a pilot for cost-sensitive agent workloads; it is not yet a proven frontier replacement, because independent scores so far run lower than DeepSeek's own numbers.
What changed vs the previous Flash
- Agentic benchmarks up across the board (vendor-reported, Hugging Face model card): Terminal-Bench 4.0 31.2 → 42.3, ProgramBench 20.3 → 29.1, DeepSWE v1.1 74.2 → 77.0
- Shorter outputs, faster task completion, per DeepSeek — which would cut cost per task beyond the token price (reasoning-trace length in early user tests is disputed; see FAQ)
- Native vision built into the model rather than bolted on
- KV cache reportedly ~1/4 the memory of the previous Flash generation (vendor claim, serving-cost relevant)
- New paid Search & Fetch APIs: $0.75 per 1,000 calls each, billing starts Oct 31, 2026
- Same serving name: the API model is still
deepseek-flash— the snapshot changed underneath a stable name, so version pinning is on you
The numbers, with trust labels
| Benchmark | Flash-1017 | Previous Flash | Source & trust |
|---|---|---|---|
| DeepSWE v1.1 | 77.0 | 74.2 | DeepSeek model card — vendor-reported, Mini-SWE-V2 harness |
| Terminal-Bench 4.0 | 42.3 | 31.2 | DeepSeek model card — vendor-reported, run with the Claude Code harness |
| HLE (no tools) | 42.6 | 36.8 | DeepSeek model card — vendor-reported |
| Intelligence Index | 39.5 | — | Artificial Analysis via OpenRouter — independent, snapshot version unconfirmed |
| Terminal-Bench (AA run) | 26.8 | — | Artificial Analysis via OpenRouter — independent, "Max" config, harness differs |
How to read this table: the vendor and independent Terminal-Bench numbers (42.3 vs 26.8) are not a contradiction you can average away — agentic scores move a lot with the harness and settings, and it is not yet confirmed the independent run used the 1017 snapshot. Treat the vendor column as "ceiling under DeepSeek's setup" and the independent column as "early outside signal". DeepSeek's own comparison tables also show Opus 5 still ahead on HLE, Terminal-Bench 4.0, ProgramBench and NL2Repo — the lead is real only on selected benchmarks.
What it costs (per 1M tokens, checked Oct 23, 2026)
| Route | Input | Output | Note |
|---|---|---|---|
| Cheapest OpenRouter provider | $0.05 | $0.60 | Provider spread is wide in launch weeks |
| DeepSeek's own endpoint | $0.15 | $0.60 | Reference price |
| Most expensive listed provider | $0.30 | $1.20 | Same model, 6× input spread — provider choice is a real decision |
Runs on what?
- API-first for almost everyone. The weights are MIT-licensed and on Hugging Face/ModelScope, but the backbone is data-centre scale (sources disagree: OpenRouter cites a 552B backbone; the HF repo reportedly lists ~763B parameters — DeepSeek has not explained the difference, and we flag it rather than pick one).
- Self-hosting friction is real: no Jinja chat template ships with the model — DeepSeek provides a Python reference encoder and a Rust library (
deepseek-recipe) instead. Community quantizations (34+ within days) lower the bar, not to laptop level. - Reported architecture: sparse MoE with asymmetric activation (~8B active on input, ~16B on generation), 1M-token context, recommended sampling temperature 1.0 / top_p 1.0 / max_tokens ≥ 256K.
Should you switch? By situation
- You run high-volume agent or coding loops on a budget → Pilot it. The cost-per-task math is the story here; run your own eval set first, because the headline scores are vendor-run.
- You are on the previous Flash in production → Upgrade in staging, pin the snapshot. Same API name, new behaviour underneath; dated snapshots are DeepSeek's cadence now, so pinning is a permanent habit, not a one-off.
- You need frontier-level hard reasoning → Stay for now. DeepSeek's own tables concede several benchmarks to Opus-class models, and independent verification is thin.
- You self-host for privacy/compliance → Evaluate, with eyes open. MIT licence and smaller KV cache help; the missing chat template and model size are your real costs.
FAQ
Is DeepSeek V4.1 Flash-1017 better than Opus 5?
On one vendor-reported benchmark, yes: DeepSWE v1.1, 77.0 vs 74.0. On HLE, Terminal-Bench 4.0, ProgramBench and NL2Repo, DeepSeek's own comparison tables put Opus 5 ahead. All figures are DeepSeek-reported and not independently verified.
Why do the benchmark numbers disagree between sources?
Agentic benchmarks depend heavily on the harness (the software running the model), settings and snapshot version. DeepSeek's Terminal-Bench run used the Claude Code harness; the independent Artificial Analysis run used a different setup and may not have tested the 1017 snapshot. Compare only like-for-like runs.
Is it really open source?
The weights are published under the MIT License — among the most permissive licences, allowing commercial use and modification. "Open weights with an MIT licence" is the precise description.
What does the "-1017" mean?
It is a dated snapshot identifier (October 17). DeepSeek ships incremental dated checkpoints instead of waiting for major versions, so the same API name can serve a newer model over time.
Any catch with the low price?
Three: provider prices vary up to 6× for the same model; early user reports mention long reasoning traces before answers, which can offset the "shorter output" claim in interactive use; and the companion Search/Fetch APIs start billing on Oct 31, 2026.
What we could not verify
- Total parameter count: 552B (OpenRouter) vs ~763B (Hugging Face listing) — unexplained discrepancy
- Whether independent Artificial Analysis runs used the 1017 snapshot
- Peak/off-peak pricing pattern reported for the Flash line in early October — recheck DeepSeek's pricing page before budgeting
Sources: DeepSeek official announcement & API docs; Hugging Face model card (figures as reported via is4.ai's Oct 18, 2026 launch analysis, checked Oct 23, 2026); OpenRouter model listing incl. Artificial Analysis figures; Local AI Zone October 2026 model tracker. Vendor-reported figures are labelled as such wherever they appear on this card.