DeepSeek V4.1 Flash 1017 gains 11 points on Terminal-Bench 4.0 under an MIT license

DeepSeek calls this a minor update. The model card says it also reduces task execution time and output length, so tasks complete faster and at lower cost, but it gives no figures for either reduction. Commenters on Hacker News pointed out that "minor update" sits oddly next to benchmarks showing substantial improvement. One commenter read it as DeepSeek continuing the 4.1 training run rather than changing how the model is trained or served. The sources describe no architecture change, and OpenRouter treats the checkpoint as a snapshot of V4.1 Flash rather than a new model.

All figures below come from DeepSeek's Hugging Face model card. Each line shows the previous V4.1 Flash score, then the 1017 score.

Opus 5 still leads the same table on several benchmarks: HLE (56.3), Terminal-Bench 4.0 (51.8), ProgramBench (37.0), NL2Repo-Bench (75.3), Chartography (84.0) and BabyVision (94.1). GPT-5.6 Sol leads ZeroBench-main at 53.0. The 1017 checkpoint leads every listed model on DeepSWE v1.1 (77.0), Automation-Bench (60.5), Agents' Last Exam (35.4) and Apex Bench (72.1). These scores are self-reported, and the harness matters. OpenRouter shows Artificial Analysis scoring "DeepSeek V4.1 Flash (Max)" at 26.8% on Terminal-Bench 4.0 and 39.2% on HLE, and the page does not say which checkpoint or harness was used, so the gap with DeepSeek's 42.3 cannot be read as a contradiction.

The same Artificial Analysis listing shows AA-Omniscience Accuracy of 46.4% and an AA-Omniscience Non-Hallucination Rate of 3.5%. The source does not define the latter metric. Teams that care about factual grounding should look it up before relying on the model's recall. OpenRouter's AutoExacto runs on five named providers put GPQA Diamond between 89.7% and 92.0% and TAU-Bench between 75.9% and 79.3%, so provider choice moved those results by up to about 3 points.

OpenRouter describes V4.1 Flash as a sparse mixture-of-experts model and the first built on DeepSeek's Causal Encoder-Decoder (CED) architecture. It activates 8B parameters on input and 16B on output from a 552B-parameter backbone. Image understanding is native, with visual and text embeddings trained jointly from the start of pre-training. In the earlier experimental V4 Flash Vision Exp, vision was added afterward. On Hugging Face, the repo is tagged image-text-to-text, and DeepSeek reports that V4.1 Flash exceeds V4 Pro on performance, speed and task completion time.

The Hugging Face safetensors metadata gives a different size: 763,205,315,794 parameters in total. By storage type, that is about 557B in I8, 204B in F8_E4M3, about 2B in BF16 and about 42M in F32. Neither source reconciles 763B with OpenRouter's 552B backbone. Treat the active-parameter figures as the ones that drive per-token compute, and the file size as the figure that drives memory.

The model card cites a technical report titled "DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression." OpenRouter says compressed KV caching cuts cache memory to roughly a quarter of the previous Flash generation. Neither source says whether "previous generation" means V4 Flash or the earlier V4.1 Flash checkpoint.

The model card states that the repository and the model weights are licensed under the MIT License. The sources do not say what license the earlier V4.1 Flash weights carried, so they cannot confirm whether this is a change. The recommended sampling settings are temperature 1.0, top_p 1.0, a 1M-token context window and max_tokens of at least 256K. Harnesses that cap generation well below that may cut off long outputs.

This release ships no Jinja-format chat template, so stacks that depend on the tokenizer's chat template have nothing to apply. DeepSeek instead provides a Python reference implementation, encoding.py, with test cases for multi-turn conversations, tool calling, thinking mode, numeric reasoning effort, mid-conversation system messages and interleaved image content. For production, DeepSeek offers deepseek-recipe, a set of Rust libraries with Python bindings. It converts Messages, Chat Completions and Responses API requests into the Conversation format, encodes them into DeepSeek V4 and V4.1 prompts or token IDs, and parses model output back into complete or streamed responses. Model inference, tool execution and HTTP transport remain the caller's job.

OpenRouter's headline price is $0.05 input and $0.60 output per million tokens, matching Relace, the cheapest posted input. Posted prices and P50 performance for selected providers:

DeepSeek's first-party endpoint posts $0.60 output, the same as Relace and Cloudflare. Its $0.003 cache-read price ties Cloudflare and Wafer for the lowest among undiscounted endpoints. At 59 tokens per second, it is slower than every standard-routing endpoint except DeepInfra Priority at 55. Baseten is about 3.4 times faster at double the output price.

OpenRouter reports 99.93% availability over three days. Over the last 24 hours, it reports 99.94% with routing and 96.83% without, which is the case for allowing provider fallback. The model is not currently available through OpenRouter's EU in-region routing. The sources give no pricing for the previous V4.1 Flash checkpoint, so they do not show whether 1017 is cheaper per token. DeepSeek's claim is that it is cheaper per task because outputs are shorter, and it publishes no figure for that.

The safetensors shards total 510,304,178,606 bytes, about 510 GB, stored largely in I8 and FP8. One commenter on Hacker News called local use tough and very costly. Another asked whether a pair of DGX Spark machines would work, putting the hardware at roughly $8k at initial retail price. The sources give no memory figures for DGX Spark, no benchmarks on it and no tested multi-node recipe. Any local setup has to hold roughly 510 GB of weights plus the KV cache for its target context, and the repo's inference folder covers weight conversion and local inference.

Hugging Face lists 42 quantizations, 5 finetunes and 1 merge in the model tree, but the sources report no sizes or quality numbers for any of them. The Hugging Face inference provider list for the repo is also empty. Until someone evaluates a quantization against the scores above, the hosted endpoints are the only options with published results in these sources.

OpenRouter says DeepSeek serves this snapshot under the deepseek-flash API model name. One Hacker News commenter noted that opencode's slug is also just deepseek-flash, which suggests the alias moves forward rather than keeping old versions in parallel. Another commenter believed OpenCode had been serving the new model for two to three days before the announcement, which the sources do not confirm. A pipeline calling the deepseek-flash alias has likely already switched to the 1017 checkpoint. To compare behavior deliberately, pin the OpenRouter slug deepseek/deepseek-v4.1-flash-1017; the sources do not say whether the previous checkpoint is still callable by its own name.

Before trusting the swap, rerun your own evals at temperature 1.0 and top_p 1.0. Measure output length and wall-clock time per task, since those are the gains DeepSeek claims without numbers. One commenter said lower thinking time and token usage is what this bump focuses on. Check that your prompt formatting matches encoding.py, and weigh the 3.5% Artificial Analysis non-hallucination figure in flows that rely on recall without retrieval. Agent builders on DeepSeek's stack can also use the new Search and Fetch APIs, priced at $0.75 per 1,000 successful requests, which take the same bearer authentication at /v1/tools/search and /v1/tools/fetch on api.deepseek.com.