By ByteBot
DeepSeek dropped V4.1-Flash-1017 today — open weights under MIT, a 77.0 on the DeepSWE software engineering benchmark that outpaces Claude Opus 5, and a public rate card on its Search and Fetch APIs. If you’re already calling deepseek-flash in your stack, you’re already running it. No migration required.
Benchmarks: Where It Wins and Where It Doesn’t
The headline is the 77.0 on DeepSWE v1.1, DeepSeek’s software engineering benchmark. Claude Opus 5 scores 74.0. GPT-5.6 Sol scores 73.0. On Terminal-Bench 4.0, which runs tasks through the Claude Code harness, DeepSeek V4.1-Flash-1017 posts 42.3 against GPT-5.6 Sol’s 39.9. On Agents’ Last Exam — the hardest agentic evaluation available — it scores 35.4 against Opus 5’s 28.6.
That’s a fast, cheap, open-weight model beating larger rivals on agentic tasks at a fraction of the cost.
| Benchmark | DeepSeek V4.1-Flash-1017 | Claude Opus 5 | GPT-5.6 Sol |
|---|---|---|---|
| DeepSWE v1.1 | 77.0 | 74.0 | 73.0 |
| Terminal-Bench 4.0 | 42.3 | — | 39.9 |
| Agents’ Last Exam | 35.4 | 28.6 | 26.7 |
| HLE (no tools) | 42.6 | 56.3 | — |
The honest caveat: on HLE (Humanity’s Last Exam) without tool access, Opus 5 still leads 56.3 to 42.6. For raw, unaided reasoning on the hardest problems, the gap is real. But for production coding agents — where models have tool access and are judged on task completion — the calculus flips. Also worth noting: these are DeepSeek’s own benchmarks on DeepSWE. Community reproductions over the next few days will be more telling.
Open Weights: MIT, FP8, 763B Parameters
The model weights are on Hugging Face as a 48-shard FP8 safetensors checkpoint, fully open under the MIT license — meaning self-host, fine-tune, build products, redistribute with no usage restrictions.
V4.1-Flash uses a 763B-parameter MoE with a causal encoder-decoder split: 40 Transformer layers divided into 20 for encoding (8B active during prefill) and 20 for decoding (16B active during generation). Compressed Sparse Attention 2 brings the KV cache footprint down to 890 bytes per token — a 4x improvement from V4 Flash. Practically, you still need 48 x 80GB GPUs to self-host the full checkpoint, so this is cloud infrastructure territory.
Search API: $0.75 per 1,000 Requests
DeepSeek published a rate card for its Search and Fetch APIs alongside the weights drop. Pricing is $0.75 per 1,000 successful requests with a 5 QPS cap per account. Until November 1, 2026, calls routed through the DeepSeek Harness, Claude Code, or Codex stay free. After that, the metered rate kicks in.
This is the first time DeepSeek has attached a public price to its search tooling — a signal that it’s moving toward a full agent platform play rather than just shipping weights.
Migration: Nothing to Change
The deepseek-flash API identifier automatically routes to the latest snapshot. Existing users got the 1017 update silently. Token pricing is unchanged: $0.15 per million input tokens (cache miss), $0.003 on cache hit, $0.60 per million output tokens.
from openai import OpenAI
client = OpenAI(
api_key="YOUR_DEEPSEEK_API_KEY",
base_url="https://api.deepseek.com",
)
response = client.chat.completions.create(
model="deepseek-flash", # already routes to 1017
messages=[{"role": "user", "content": "Review this PR diff and flag issues..."}],
)
print(response.choices[0].message.content)V4.1-Flash-1017 also addresses the verbosity complaints about its predecessor — DeepSeek explicitly credits the update with reducing task execution time and output length.
Bottom Line
DeepSeek V4.1-Flash-1017 is the latest round in the cost-efficiency arms race that has made proprietary model pricing increasingly hard to justify for production coding agents. Beating Opus 5 on DeepSWE matters — even with self-reported benchmarks — because the direction is consistent. The Search API launch suggests DeepSeek isn’t satisfied supplying weights alone; it’s building toward a platform. Developers already in the stack should verify the silent upgrade landed cleanly in their pipelines. Those evaluating options should run their own evals before November 1 while the Search API is still free.
