THE SOURCE IN CONTEXT
The 17 October checkpoint: official weights, protocol details and minimal local inference. Start here to identify 1017 before reading a deployment recipe.
What changed in the 1017 checkpoint?
DeepSeek published this weight refresh on 17 October 2026. The model name includes 1017, so it identifies a different artifact from the September launch checkpoint used by the older workstation reports. Start from the pinned weight revision on this page and preserve the encoder, tokenizer and conversion identity with it. A runtime recipe that now defaults to 1017 does not retroactively update an earlier speed experiment.
Source: deepseek-ai · pinned document
How should prompts and tool calls be encoded?
The release supplies a Python encoding reference instead of a Jinja chat template. Its maintained deepseek-recipe toolkit translates API messages into the model protocol and parses responses. That toolkit does not perform inference, run tools or provide HTTP transport. A generic chat-template substitution can therefore change behavior even when the weights load successfully.
Source: deepseek-ai · pinned document
What do the capability figures establish?
The publisher reports HLE without tools as 42.6 on the full benchmark and 44.0 on its text-only subset. These populations differ. DeepSWE v1.1 uses Mini-SWE-V2, while Terminal-Bench 4.0 uses Claude Code. The figures describe the publisher’s evaluation protocol; no local GPU configuration or tokens-per-second measurement is established by that table.
Source: deepseek-ai · pinned document
Which inference settings should travel with a reproduction?
The card recommends temperature 1.0, top_p 1.0, a 1M-token context and an output budget of at least 256K tokens. These are publisher recommendations, not evidence that a particular workstation can allocate that window. The separate minimal-inference example is a reference implementation requiring conversion into per-rank weights; it is not a drop-in production server.
Source: deepseek-ai · pinned document
What does the source specify?
These declarations belong to the cited document. A planning minimum, model limit and measured runtime allocation are different quantities.
| Recommended sampling | temperature 1.0; top_p 1.0 |
|---|---|
| Published context / recommended output budget | 1M context; max_tokens at least 256K |
| Prompt interface | Python encoder; no Jinja-format chat template |
Reported figures and their conditions
Selected facts reported by the original author. Each row has its own configuration and measurement boundary. Run AI Here has not reproduced these tests.
| Configuration / workload | Metric | Reported value | Evidence |
|---|---|---|---|
| HLE without tools, full set | HLE | 42.6 reported scorereported value | Original row ↗ |
| HLE without tools, text-only subset | HLE | 44.0 reported scorereported value | Original row ↗ |
| DeepSWE v1.1 | DeepSWE | 77.0 reported scorereported value | Original row ↗ |
| Terminal-Bench 4.0 | Terminal Bench | 42.3 reported scorereported value | Original row ↗ |
HLE without tools, full set · HLE · conditions
| model | deepseek-ai/DeepSeek-V4.1-Flash-1017 |
|---|---|
| model_revision | 1bae0e1088993b9b82de08f59de5be5a34e42170 |
| hardware | Not specified |
| executed_runtime_revision | Not specified |
| input_tokens | Not specified |
| output_tokens | Not specified |
| repetitions | Not specified |
| measurement_scope | publisher capability evaluation, not local speed |
| benchmark | HLE without tools |
| population | full set |
| tools | False |
HLE without tools, text-only subset · HLE · conditions
| model | deepseek-ai/DeepSeek-V4.1-Flash-1017 |
|---|---|
| model_revision | 1bae0e1088993b9b82de08f59de5be5a34e42170 |
| hardware | Not specified |
| executed_runtime_revision | Not specified |
| input_tokens | Not specified |
| output_tokens | Not specified |
| repetitions | Not specified |
| measurement_scope | publisher capability evaluation, not local speed |
| benchmark | HLE without tools |
| population | text-only subset |
| tools | False |
DeepSWE v1.1 · DeepSWE · conditions
| model | deepseek-ai/DeepSeek-V4.1-Flash-1017 |
|---|---|
| model_revision | 1bae0e1088993b9b82de08f59de5be5a34e42170 |
| hardware | Not specified |
| executed_runtime_revision | Not specified |
| input_tokens | Not specified |
| output_tokens | Not specified |
| repetitions | Not specified |
| measurement_scope | publisher capability evaluation, not local speed |
| benchmark | DeepSWE v1.1 |
| harness | Mini-SWE-V2 |
Terminal-Bench 4.0 · Terminal Bench · conditions
| model | deepseek-ai/DeepSeek-V4.1-Flash-1017 |
|---|---|
| model_revision | 1bae0e1088993b9b82de08f59de5be5a34e42170 |
| hardware | Not specified |
| executed_runtime_revision | Not specified |
| input_tokens | Not specified |
| output_tokens | Not specified |
| repetitions | Not specified |
| measurement_scope | publisher capability evaluation, not local speed |
| benchmark | Terminal-Bench 4.0 |
| harness | Claude Code |
AI-assisted source reading and original editorial synthesis; no independent model execution.