DeepSeek-V4-Flash Update
DeepSeek has released a major upgrade to its V4-Flash language model, reporting large gains on coding and tool-use benchmarks that put it in the same league as leading proprietary models like OpenAI’s latest GPT-5.6 variants—while remaining dramatically cheaper per token, especially when cache hits are leveraged. Commenters highlight that the open weights, relatively modest hardware requirements, and strong price‑performance make it attractive for local inference, coding agents, and high‑volume “grunt work” tasks, even if it still lags top models on some reasoning and multimodal capabilities. Alongside enthusiasm, there are concerns about opaque benchmarking, data privacy and Chinese jurisdiction, and the operational risks of silently upgrading a model behind a stable API name.
Model performance & benchmarks
- Update shows large jumps vs prior V4-Flash on internal harness: big gains on Terminal Bench and Toolathlon.
- Compared to GPT‑5.6 Terra/Luna: Flash now beats them on some coding/terminal benchmarks (Terminal Bench, Toolathlon, CyberGym) but lags on others (DeepSWE, Agents’ Last Exam). No clear overall winner.
- Some point to official leaderboards (e.g., Terminal-Bench, DeepSWE) as the “real” scores; others note labs’ self-reported numbers and possible typos.
- Multiple users report Flash feeling close to frontier models (Opus 4.7–5.x, Sonnet 5, Luna) in coding tasks at much lower cost.
Pricing, caching & cost-per-task
- V4-Flash is repeatedly described as “insanely cheap,” especially with caching.
- Shared usage stats: hundreds of millions to ~2 billion tokens for under $20 in a couple weeks, thanks to ~99% cache hit rates.
- Direct DeepSeek API is seen as cheaper and better for caching than some aggregators; others use third‑party plans (e.g., OpenCode Go) that resell DeepSeek with ZDR‑like guarantees, though recent wording changes cause uncertainty.
- Comparisons with Luna highlight Flash cache read at $0.0028/MTok vs Luna at $0.02/MTok.
Local inference & hardware
- Model: DeepSeek‑V4‑Flash‑284B‑A13B, ~160 GiB full weights, described as “only 160 GiB” for a ~200–300B MoE‑class model.
- Runs well on 2× DGX Spark or 2× RTX 6000/6000 Pro; single B300 is borderline; 128 GB M5 Max can run strong quantizations but slowly.
- 4‑bit and even 2–3‑bit quantizations discussed; community projects (e.g., ds4) are rapidly updating.
Use cases & harnesses
- Heavy use for coding: C#, Rust, C/C++, Go, Python, TS/JS, SQL, CSS/HTML; especially with agent harnesses like pi, OpenCode, and other subagent systems.
- Users emphasize that “harness matters more than model” for productivity: subagents, MCP tools, caching, context reducers, and custom “skills” for code review.
- Also used for QA on voice transcripts, classification, extraction, rewriting, and deep web research via search MCPs.
- Some pair a “frontier” planner (Opus/Sol/Luna/K3) with V4-Flash as the cheap executor.
Open weights, data & governance
- Weights are released on Hugging Face; many see this as key for long‑term “usable forever” capabilities and better data protection via self‑hosting.
- Concerns about sending code to Chinese‑hosted APIs; mitigations include using ZDR‑style intermediaries or local deployment.
- Debate over future of Chinese open weights and potential US regulation; participants cite conflicting readings of political speeches and intentions.
Versioning & evaluation skepticism
- Confusion over model naming: old and new Flash both exposed as
deepseek-v4-flashon DeepSeek; some providers append-0731. - Several argue this should have been “V4.1” to avoid silent behavioral changes in production.
- Some are wary of benchmark‑driven marketing and suspect fine‑tuning to benchmarks; a few say they largely ignore benchmarks and rely on real‑world performance instead.