DeepSeek V4.1 Flash vs V4 Flash in Harness
DeepSeek V4.1 Flash vs V4 Flash in Harness
- Author
- DeepSeekAgent.io Editorial Team
- Published
- Updated
DeepSeek V4.1 Flash replaces the original V4 Flash API route and changes the model architecture, multimodal support, cache footprint, and reported Agent performance. For Harness users, the key question is operational: does the new model finish tool-heavy work more reliably or efficiently under the same scaffold?
Official evaluations provide a strong first comparison. A community quick test adds real DSH runtime data, but its unfinished task state prevents a clean quality conclusion. This article keeps those two evidence levels separate.
Release-level differences
| Dimension | DeepSeek V4 Flash | DeepSeek V4.1 Flash |
|---|---|---|
| API status | Retired ID, temporarily aliased | Current deepseek-flash route |
| Backbone parameters | 284B | 552B |
| Active parameters | 13B | 8B prefill / 16B decode |
| Native image input | Separate Vision Exp route | Included in the current model |
| Context | 1M in the official comparison | 1M |
| Global KV cache | Baseline | 890 bytes/token, about one quarter of V4 Flash |
| Reasoning control | Earlier generation | Continuous 1–100 at API level |
The asymmetric 8B/16B activation matters for Agent workloads because their input and output shapes differ. A long repository context and tool history can dominate prefill, while each action may generate much less text. The Causal Encoder–Decoder architecture reduces work on the input-heavy side and uses a larger active path during generation.
Official same-family comparison
DeepSeek reports the following instruct results at maximum reasoning effort, temperature=1.0, and top_p=0.95. Code Agent evaluations use a one-million-token context and named Harnesses.
| Benchmark | V4 Flash | V4.1 Flash | Change |
|---|---|---|---|
| Terminal-Bench 2.1 | 82.7 | 90.6 | +7.9 |
| Terminal-Bench 3.0 | 7.6 | 30.0 | +22.4 |
| Terminal-Bench 4.0 | 7.0 | 31.2 | +24.2 |
| DeepSWE v1.1 | 54.4 | 74.2 | +19.8 |
| NL2Repo-Bench | 54.2 | 64.0 | +9.8 |
| CyberGym | 76.7 | 88.1 | +11.4 |
| SEC-Bench Pro | 30.9 | 62.8 | +31.9 |
These figures support a meaningful V4.1 improvement within DeepSeek's evaluation. They do not isolate architecture alone: DeepSeek also changed pretraining, the Agent-task data pipeline, reinforcement-learning scale, and on-policy distillation data.
One methodology detail deserves emphasis. DeepSWE uses the official mini-SWE harness in the primary model comparison, while Terminal-Bench uses DSH Minimal. SEC-Bench Pro uses Claude Code. The score belongs to a model-and-scaffold configuration.
What the DSH community quick test found
A September 2026 Reddit post ran V4.1 Flash and V4 Flash Vision Exp in the latest DeepSeek Harness with the same app-building prompt, Skill file, and specification. The author reported:
| Reported metric | V4.1 Flash | V4 Flash Vision Exp |
|---|---|---|
| Wall time | 18m 49s | 30m 11s |
| Total usage | 11.59M | 20.31M |
| Output | 126K | 154K |
| LLM time | 7m 48s | 25m 10s |
| Generation speed | 361 tok/s | 120 tok/s |
| Cache hit | 99.5% | 99.7% |
For that run, V4.1 finished about 38% sooner, reported about 43% fewer total tokens, and showed roughly three times the generation speed.
The screenshot also showed one task still in progress and one pending for V4.1, while the Vision Exp run had completed all 11. The outputs therefore differ in completion state. Runtime and token reductions are useful observations, but they cannot establish equal-quality completion or a universal percentage advantage.
How to run a cleaner Harness comparison
Use one controlled task suite and publish the full setup:
- Pin the DSH version and Profile.
- Use the same official endpoint, region, machine, tools, permissions, and Skill files.
- Start fresh Sessions with equivalent context.
- Fix reasoning effort, temperature, output cap, and step cap.
- Verify that both runs finish every required subtask before comparing time.
- Repeat each task and report variance, retries, cache hits, billed tokens, and actual cost.
- Score the artifact with tests or a blinded rubric, not only visual impressions.
- Save trajectories and classify model, tool, Harness, and environment failures separately.
Run DSH Minimal and Standard independently when the choice matters. DeepSeek's own V4.1 scaffold table reports 72.6 versus 70.5 on DeepSWE and 90.6 versus 85.8 on Terminal-Bench 2.1, showing that one Profile cannot represent all DSH behavior.
Should existing DSH users migrate?
For new official-API work, deepseek-flash is the current route and the old Flash IDs are retired compatibility aliases. Migration is the sensible default after a smoke test.
Keep a short rollback comparison for critical workflows. Test your longest Sessions, image attachments, repository navigation, tool-call recovery, and acceptance tests. One community report describes a reasoning loop during an exceptionally long maximum-effort V4.1 session. That single report should guide diagnostics, while broader reliability claims require more data.
The practical target is a measured task result: correctness, elapsed time, cost, trajectory quality, and recovery behavior together.
Related reading
- DeepSeek V4.1 Flash overview
- Configure V4.1 Flash in DeepSeek Harness
- Reproduce DeepSWE with DSH Minimal
- Historical DeepSeek V4 Flash page
FAQ
Is V4.1 Flash always faster than V4 Flash in DSH?
No universal percentage has been established. Official architecture and benchmark results point to substantial gains, and one community run reports a large speedup, but workload, cache state, completion quality, endpoint load, and DSH configuration all matter.
Does the community test prove 43 percent lower token use?
It reports that reduction for one run. The V4.1 screenshot still had unfinished tasks, so the token totals do not represent equal completed work.
Can I continue using the old model ID?
DeepSeek says the retired V4 Flash IDs temporarily route to V4.1 Flash. Use deepseek-flash for new configurations and treat the aliases as a migration window.