DeepSeek-V4.1-Flash-1017 ships: free MIT update, faster reasoning, shorter outputs
DeepSeek shipped the 1017 checkpoint of V4.1-Flash on Oct 17 (MIT). Broad official gains: HLE 36.8→42.6, Terminal-Bench 31.2→42.3, Apex 64.6→72.1; shorter outputs mean faster wall-clock and lower cost. Local quants landed same-week; catalog row added on our site (763B total stays catalog-only, no single-GPU runs).
Source: araoai
DeepSeek-V4.1-Flash-1017 ships: free MIT update, faster reasoning, shorter outputs
On Oct 17, DeepSeek published DeepSeek-V4.1-Flash-1017 on Hugging Face (MIT) — a dated checkpoint refresh of the original DeepSeek-V4.1-Flash. Architecture and deployment path stay the same (763.2B total / 37B active MoE, 1M context); what changed is weight quality and output behavior.
Benchmark gains (official numbers)
| Benchmark | V4.1-Flash | 1017 | Δ |
|---|---|---|---|
| HLE (w/o tools) | 36.8 | 42.6 | +5.8 |
| HLE text-only (†) | 39.1 | 44.0 | +4.9 |
| Terminal-Bench 4.0 | 31.2 | 42.3 | +11.1 |
| DeepSWE v1.1 | 74.2 | 77.0 | +2.8 |
| ProgramBench | 20.3 | 29.1 | +8.8 |
| NL2Repo-Bench | 65.4 | 68.2 | +2.8 |
| Automation-Bench | 54.8 | 60.5 | +5.7 |
| Agents Last Exam | 31.8 | 35.4 | +3.6 |
| Apex Bench | 64.6 | 72.1 | +7.5 |
| Chartography | 78.9 | 83.0 | +4.1 |
| BabyVision | 89.6 | 92.5 | +2.9 |
The biggest jumps are agentic workloads (Terminal-Bench +11.1, ProgramBench +8.8), with general reasoning close behind (HLE +5.8, Apex +7.5).
Shorter outputs = faster and cheaper in practice
DeepSeek explicitly markets shorter outputs: fewer generated tokens per task means lower wall-clock latency and lower API bills. Combined with the score gains, 1017 is a free update that does more for the same money.
Deployment notes
- No Jinja chat template: conversation formatting moved to the official deepseek-recipe (Rust library); community inference stacks need to follow
- 1M context, max_tokens officially recommended at 256K or above
- Local quants spread within two days: GGUF, mxfp4, EXL3 and GSQ-agentic-strata are all out
How we logged it
Our model catalog now lists DeepSeek-V4.1-Flash-1017 (shares the 37B active-params convention with the original, MIT → open source). The 763B total keeps it out of single-GPU benchmark scope — the original V4.1-Flash has the same catalog-only, zero-records stance and this batch does not break it. Running it locally is waiting on multi-GPU or expert-offload setups, not on a bigger single card.
Context
The official comparison column also lists K3, GLM-5.3, GPT-5.6 Sol and Opus 5. Press coverage focused on two angles: the free, MIT-licensed update itself, and how shorter outputs pressure the inference-chip demand narrative (chip stocks slipped on release day).