DeepSeek's Flash Refresh Beats Its Predecessor, Still Trails Opus 5
DeepSeek's Flash update leads its predecessor on every benchmark in its model card and undercuts rivals on price and latency, though Opus 5 still tops five of them.
DeepSeek released V4.1-Flash-1017, a minor update to V4.1-Flash that the model card says cuts task execution time and output length while improving reasoning, coding, agentic and visual benchmarks. The 763B-parameter model is MIT-licensed and ships a Python prompt-encoding reference plus Rust libraries instead of a Jinja chat template.
On the card's own numbers the update leads its predecessor on every listed benchmark, including Terminal-Bench 4.0 at 42.3 versus 31.2 and Apex Bench v1.1 at 72.1 versus 64.6, while trailing Opus 5 on HLE, Terminal-Bench, ProgramBench, NL2Repo-Bench and BabyVision. Terminal-Bench and DeepSWE runs used the Claude Code and Mini-SWE-V2 harnesses.
Key takeaways
- Sampling defaults are temperature 1.0 and top_p 1.0, with a 1M-token context window and max_tokens of at least 256K. The repository lists 763B parameters in BF16, F32, F8_E4M3 and I8 tensor types, 10,233 downloads in the last month, and an MIT license covering weights and code.
- Terminal-Bench 4.0 and DeepSWE v1.1 were run with the Claude Code and Mini-SWE-V2 harnesses respectively; the HLE figure of 42.6 rises to 44.0 on the text-only subset, marked with a dagger.
- Opus 5 leads the comparison table on HLE (56.3), Terminal-Bench 4.0 (51.8), ProgramBench (37.0), NL2Repo-Bench (75.3) and BabyVision (94.1), against 42.6, 42.3, 29.1, 68.2 and 92.5 for the new model.
- DeepSeek leads on Apex Bench v1.1 (72.1 versus 65.8), Agents' Last Exam (35.4 versus 28.6), Automation-Bench v1.0.6 (60.5 versus 50.3) and DeepSWE v1.1 (77.0 versus 74.0).
- No Jinja chat template is included. The encoding folder holds a Python reference implementation with test cases for multi-turn conversations, tool calling, thinking mode, numeric reasoning effort, mid-conversation system messages and interleaved images; deepseek-recipe adds Rust libraries with Python bindings that convert Messages, Chat Completions and Responses API requests, leaving inference, tool execution and HTTP transport to the caller.
What people think
Analysis details
36 comments analyzed. Analyzed .
The sample selects active discussion branches and includes available parent comments. It may omit parts of the full thread. The analysis included 36 of 103 usable stored comments because of the input limit. Selected themes do not measure community opinion or how common a view is.
Cheap mid-tier work and agentic coding
DeepSeek and Muse are named as go-to choices for basic and mid-tier tasks, with opencode described as a good host. The counterweight is scope: for the large majority of tasks any recent model would do, and gaps between frontier models show up mainly in long-running or very difficult work, where they are small. The practical case rests on price and latency rather than benchmark leadership.
Subscription burn and cheap per-token inference
One workflow exhausts a $200 Claude Code subscription in under 24 hours on a single project, spawning subagents across a homelab rack; included API credits prompted a rewrite of the coding agent setup. Cheap per-token pricing and near-free cache reads offset DeepSeek's longer thinking, since higher throughput absorbs the extra tokens.
Overthinking, thinking budgets and quantization
Earlier releases drew complaints for overthinking, questioning themselves instead of starting, and the suggested fix is to run high or low reasoning effort rather than maximum for implementation. Visible chain-of-thought is not assumed to reflect the computation behind an answer, with interpretability papers cited. An open question asks which quantization an external host serves.
Retention terms, routing and cyber-review blocks
DeepSeek's cloud terms allow submitted content to be used for training, while Fireworks and comparable hosts offer zero-data retention; opencode routes DeepSeek requests to China. Separately, sessions can halt when a prompt trips cybersecurity review, a failure mode reported as worst with one rival. Whether third-party hosts of Chinese models inherit the same restrictions is left open.
Rival hosts, GLM and a Z.ai mix-up
Fireworks is recommended as a non-Chinese host at an acceptable price, and OpenCode Go is noted as bundling DeepSeek models. A stated preference for DeepSeek alongside a plan to subscribe to Z.ai drew corrections that the two are different companies; Z.ai does resell DeepSeek access. GLM-5.3-flash is preferred in one case, which still credits DeepSeek with avoiding over-clever solutions.
Regulatory capture and ecosystem diversity
The release is read as evidence for why OpenAI and Anthropic seek regulatory hurdles, with ecosystem diversity framed as the durable answer to hard problems. The objection: no proposed rule would have stopped this model, which is not frontier. The rebuttal chains public warnings about dangerous open weights with advocacy for stronger government oversight into an intent to regulate them.