On Oct 17 DeepSeek upgraded V4.1 Flash to V4.1-Flash-1017 across web, app and API (deepseek-flash, same call and price) and published MIT-licensed weights on Hugging Face and ModelScope. Vendor-reported DeepSWE v1.1 rises from 74.2 to 77.0 and Terminal-Bench 4.0 from 31.2 to 42.3, with shorter runtimes and outputs. It also opened Search/Fetch APIs backed by its in-house search engine at $0.75 per 1,000 calls.
Key Takeaways
- ✓Vendor-reported DeepSWE v1.1 77.0 (prev. 74.2), above Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0) in DeepSeek's table
- ✓Terminal-Bench 4.0 31.2 → 42.3 and ProgramBench 20.3 → 29.1, still behind Opus 5 (51.8 / 37.0); HLE w/o tools 42.6 vs Opus 5's 56.3
- ✓Pricing unchanged: deepseek-flash peak $0.30 input (cache miss) / $1.20 output per 1M tokens, off-peak half; 1M context, 384K max output
- ✓Search / Fetch APIs at $0.75 per 1,000 calls each; server-side calls via DeepSeek Harness, Claude Code and Codex are free until Oct 31, 2026 16:00 UTC
- ✓MIT weights on Hugging Face in FP8 (FP4 experts); no Jinja template — use the bundled encoding.py or deepseek-recipe
Key Decision Metrics at a Glance
Heavy Claude Code use: compare subscription limits and API bills
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background
DeepSeek V4.1 Flash (Sep 10) was a cheap long-context model that trailed closed frontier models on long-horizon coding and terminal work. On Oct 17 DeepSeek replaced it with V4.1-Flash-1017 everywhere at once — web, app and API — keeping the deepseek-flash model name, the call format and the price. The retired names deepseek-v4-flash and deepseek-v4-flash-vision-exp are now served by the new model at Flash pricing.
What changed
The model card calls it a minor update to V4.1-Flash. That architecture, per the V4.1-Flash card, is a 552B-backbone multimodal MoE with a causal encoder-decoder layout that activates 8B parameters per token in prefill and 16B in decode, with a 1M context. DeepSeek says 1017 adds RL with self-feedback to improve from-scratch project building (games, 3D scenes) and also cuts task runtime and output length — a performance–cost Pareto improvement across several agent harnesses, with specific tuning for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy.
Benchmarks (vendor-reported)
DeepSWE v1.1 77.0 (from 74.2, Mini-SWE-V2 harness), Terminal-Bench 4.0 42.3 (from 31.2, Claude Code harness), ProgramBench 29.1 (from 20.3), Automation-Bench v1.0.6 60.5, Agents' Last Exam 35.4, Apex Bench v1.1 72.1, HLE w/o tools 42.6. In DeepSeek's own table it beats Kimi K3, GLM-5.3 and GPT-5.6 Sol on most agentic rows, but still trails Claude Opus 5 on Terminal-Bench 4.0 (51.8), ProgramBench (37.0), NL2Repo (75.3) and HLE (56.3). No independent reproductions yet.
Getting started
API users need no code change; peak pricing stays at $0.30/1M input (cache miss) and $1.20/1M output, half off-peak, with 1M context and 384K max output. DeepSeek Harness users should move to v0.2.1-rc.1 or later. For self-hosting, MIT weights (FP8 with FP4 experts) are on Hugging Face and ModelScope; there is no Jinja chat template, so use the bundled encoding.py or the deepseek-recipe Rust/Python toolkit. Separately, DeepSeek opened Search (POST /v1/tools/search) and Fetch (POST /v1/tools/fetch) APIs at $0.75 per 1,000 calls; server-side tool calls through Harness, Claude Code and Codex become billable from Oct 31, 2026 16:00 UTC.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.