DeepSeek V4.1 Flash Arrives on Novita AI With Image Input and Stronger Coding

What DeepSeek V4.1 Flash offers

According to Novita AI, DeepSeek V4.1 Flash is the successor to DeepSeek V4 Flash, released on September 10, 2026 under the MIT license. It is a multimodal Mixture-of-Experts model with 552B backbone parameters plus 196B parameters in an "Engram" conditional-memory module, for roughly 748B total parameters.

Unlike the text-only V4 Flash, the new model accepts text and image input and returns text output. Novita notes that developers can send a screenshot, chart or document page together with an instruction, instead of running a separate OCR or vision step first. Previously, image-aware DeepSeek workloads needed the experimental Vision variant, which Novita describes as now superseded.

The context window is 1,048,576 tokens, with a maximum output of 393,216 tokens. DeepSeek also supports a reasoning effort setting, an integer from 1 to 100, that trades inference cost for accuracy on a single model ID.

Benchmark gains and architecture changes

Novita cites DeepSeek's own model card, at maximum reasoning effort, for the following comparisons between V4.1 Flash and V4 Flash:

• Terminal-Bench 2.1 (Pass@1): 90.6 versus 82.7
• DeepSWE v1.1 (Resolved): 74.2 versus 54.4
• AutomationBench (Pass@1): 54.8 versus 37.7

These are DeepSeek's own evaluation numbers, and Novita advises validating them against your own agent harness and task set before treating them as a production guarantee.

On the architecture side, V4.1 Flash uses a Causal Encoder-Decoder design: a 40-layer transformer split into a 20-layer causal encoder and a 20-layer decoder. The decoder's KV cache is projected from the encoder's final hidden states rather than computed fresh at every layer. Combined with FP4 KV caching and a technique DeepSeek calls Compressed Sparse Attention 2, this cuts the global KV cache footprint to roughly a quarter of V4 Flash's, according to the source. The model activates 8B parameters during prefill and 16B during decode, versus a flat 13B for V4 Flash.

Novita stresses that this is an efficiency gain on DeepSeek's serving side, not a price cut. The per-token price on Novita is set by Novita's pricing.

Pricing and API access

Novita lists the model at $0.3 per 1M input tokens, $1.2 per 1M output tokens and $0.006 per 1M tokens for cache reads. That sits between V4 Flash at $0.14 and $0.28, and the V4 Flash 0731 revision at $0.44 and $1.32.

The model ID is deepseek/deepseek-v4.1-flash, and the base URL is https://api.novita.ai/v3/openai. It is available through chat completions, Anthropic-compatible and Responses endpoint families, so existing Novita integrations can switch by changing the model ID. Text-only requests should work unchanged, while image requests need multimodal message content as described in Novita's API reference.

Who should consider it, and who should not

Novita recommends V4.1 Flash for coding agents, terminal automation, multi-step tool use and browser automation, as well as long-context agent sessions and requests that mix images with text. Teams on V4 Flash are encouraged to re-test their prompts to see whether the price step-up pays off.

The post advises against defaulting to the new model for simple, high-volume, latency-sensitive text requests. It says DeepSeek's base-model benchmarks show V4.1 Flash roughly matching or trailing V4 Flash on tests such as MGSM and MATH. For short-prompt classification or chat without agentic or multimodal needs, the cheaper V4 Flash may be the better default. Because V4.1 Flash is a recent release, Novita also suggests validating tool-call schemas, image-input paths and failure modes before routing production traffic, and notes that DeepSeek Flash pricing and model status have changed more than once this year.