DeepSeek V4.1 Flash hands-on tests and comparisons

The author reads the vendor table: DeepSWE 77.0 is close to Gemini 4 Argon's 77.9 and GPT-6.1 Sol's 75.2, while HLE 42.6 ties GLM-5.3 and Qwen3.8 Max but trails four overseas models by 7 to 15 points.

The post notes V4.1 Flash 1017 is MIT open weights with 1M context, self-reporting DeepSWE 77.0 over Opus 5's 74.0 but Terminal-Bench 42.3 under 51.8, with $0.15/1M off-peak input suited to bulk work.

A user ran DeepSeek V4.1 Flash on two DGX Sparks, hitting 2,135 tok/s on a 93K-token prompt and loading 1M context in 27s, far above the old four-Spark vLLM setup's 48.5 tok/s.

A user had DeepSeek V4.1 Flash generate a Three.js 3D BMW M5 from a single prompt, costing $0.38 and taking 1 hour 10 minutes, which they found pricey and too slow.

The post says DeepSeek released V4.1-Flash-1017 on 17 October with MIT open weights, 763B parameters and 1M context, self-reporting DeepSWE 77.0 versus Opus 5's 74.0, and links the Hugging Face weights.

A test reports V4.1 Flash 1017 matches Opus 5.5 on long-horizon browser tasks while costing 19x less, noting Playwright tests need updated specs to avoid reward hacking.

A roundup of official and Hugging Face details: 763B total parameters, 1M context, MIT license, $0.15/M input off-peak, and a caveat that all benchmarks are DeepSeek-run.

The author notes the update was billed as minor yet all 11 published benchmarks improved, with Terminal-Bench 4.0 up 11.1 points, and suggests re-testing automations before switching.

A user asks why DeepSeek V4.1 Flash-1017 can be so cheap if it really distilled GPT-6 Astra, speculating about the source of its low cost.

A user says Sol suffices as a worker but V4.1 Flash 1017's low price is tempting, while wishing for an orchestrator model that rivals Opus and could run on a MacBook Pro.

A user compared DeepSeek V4.1 Flash in a single bare API call (under 100K tokens) against a harness run of 3 rounds and ~400 steps (~90M tokens, 99% cached), finding the bare call already shows model capability.

In one ThreeJS soft-body file, V4.1 Flash 1017 gets spring-damper physics and collision recovery right, while GLM 5.3 Flash EXL3 shows cleaner geometry but drift in vertex deformation.