DeepSeek V4.1 Flash hands-on tests and comparisons

Relaying Yuchen Jin: DeepSeek-V4.1-Flash-1017 scores 77 on DeepSWE v1.1, reaching Opus 5-level coding at 83x lower cost, with a Databricks launch planned.

Ollama Cloud hosts V4.1 Flash 1017 with US-based hosting and zero data retention at $0.15/$0.60 per 1M tokens off-peak, launchable in Claude Code or OpenCode with one command.

Claims V4.1-Flash-1017 tops GPT 5.6 Sol on Terminal-Bench while costing ~17x less, a 35% gain over the prior V4.1-Flash from RL on multiple coding agent harnesses rather than algorithmic novelty.

Local deployment test quantizes only routed experts to EXL3 2.9 and keeps official FP8 elsewhere, hitting 180 tok/s structured, 137 code, 97 prose, a 1.53M KV pool and 9/9 needle retrieval at 30K-241K.

After a two-year DeepSeek timeline, argues 1017 leads Opus 5 on agentic coding benchmarks like DeepSWE 77.0 but trails on HLE (42.6 vs 56.3) and Terminal-Bench 4.0 (42.3 vs 51.8), against Opus 5 rather than 5.5.

A news aggregator relays the official claim that V4.1-Flash-1017 improves reasoning, coding and agent tasks with better cost efficiency, without citing specific scores or prices.

A long architecture report compares V4.1-Flash-1017's 552B sparse MoE (8B active in prefill, 16B in decode) with KUT Gemini Flash's manifold-constrained approach, focusing on long-context reasoning efficiency and KV compression.

Vendor-reported scores show Terminal-Bench 4.0 rising from 31.2 to 42.3 and DeepSWE at 77.0 versus Opus 5's 74.0, while tool-free HLE lags at 42.6; FP4 KV cache plus CSA2 cuts KV to about 890 bytes per token.

A discussion post asks why V4.1 Flash 1017 can stay so cheap if it really distilled GPT-6 Astra, speculating about cost and training provenance.

The author reads Flash as the high-volume cheap tier, notes the date-in-name signals fast iteration, and holds off judging capability until pricing and evals appear.

The post reports DeepSeek V4.1-Flash-1017 scoring 77.0 on DeepSWE, ahead of Claude Opus 5 at far lower cost, alongside same-week open-weight releases from Reflection and StepFun.

A user used DeepSeek-V4.1-Flash with a detailed Blender prompt to produce a 15-second rainy cyberpunk street close-up preview, sharing the full prompt covering materials, wetness layers, lighting and render settings.