Changelog

New models, price changes and retirements on Entrim. Every announcement we email, in one place.

DeepSeek V4.1 Flash is retiring today

A week ago we said V4.1 Flash had a week left. That week is up, and we're retiring the model today. The DeepSeek V4.1 Flash 1017 build takes its place.

1017 is DeepSeek's October update, and it keeps everything V4.1 Flash does while getting there faster. It scores higher on agentic coding, 42.3 on Terminal-Bench 4.0 against 31.2, and it writes shorter outputs getting there, so the same job costs you less at the same price per token. The architecture and the prompt format are unchanged: same 1M-token context, same image input, same reasoning dial from 1 to 100.

One line to change

Prices are unchanged, and the endpoint is the same. The model name is the only thing that changes.

Model name
Old: "deepseek-ai/DeepSeek-V4.1-Flash"
New: "deepseek-ai/DeepSeek-V4.1-Flash-1017"

The old name will not silently redirect

We're not quietly redirecting deepseek-ai/DeepSeek-V4.1-Flash to 1017. A call to the old name won't land on the new build behind your back, it stops there.

If you hit anything unexpected on 1017, post in our Discord. Your feedback goes straight to our engineers.

Try DeepSeek V4.1 Flash 1017

DeepSeek V4.1 Flash build 1017 is available

We've updated DeepSeek V4.1 Flash to the newer 1017 build. The gains land in the long agentic runs: terminal work, program synthesis, and multi-step automation.

BenchmarkV4.11017Δ
Terminal-Bench 4.031.242.3+11.1
ProgramBench20.329.1+8.8
Apex Bench v1.164.672.1+7.5
Automation-Bench 1.0.654.860.5+5.7
HLE (no tools)36.842.6+5.8

Call it today

Build 1017 is on the API now, under its own model name:

Model name
"deepseek/deepseek-v4.1-flash-1017"

Run it against your own workload and tell us how it behaves. If you see unexpected responses, errors, or performance issues, post in our Discord. Your feedback goes straight to our engineers.

The current build will be retired on 29 October

DeepSeek V4.1 Flash will remain available via the API for another week. During this transition, you may experience slower performance than usual, as we've reallocated serving capacity to the new 1017 build.

We recommend testing your workloads on the new version before the transition is complete to identify any potential changes in model behavior and avoid unexpected issues.

#01AVAILABLEQWEN

Qwen3 Embedding 8B is on the API

We've added capacity, and Qwen3 Embedding 8B is available on our model APIs.

It's the retrieval half of a RAG stack. It took the top spot on the MTEB multilingual leaderboard at release and scores 81.22 on MTEB-Code, so it holds up on natural-language queries over repositories and docs as well as it does over prose. One embedding space covers 100+ natural and programming languages, which means you can query in German and match English documents without a second model.

A one-line task instruction steers the same vectors toward retrieval, classification, or clustering, with no fine-tuning. Vectors come out at 4096 dimensions and truncate down to 32, so you can size them to what your store can afford.

Input / M
$0.01
Context
32k
Quantization
BF16
Model name
"Qwen/Qwen3-Embedding-8B"

Full specs are on the model page


#01NEWGLM

GLM 5.3 is live

As promised, it's here. GLM 5.3 is our flagship for long engineering work: multi-hour refactors and migrations, whole-repository analysis, security review, and agents that chain tool calls for hours without losing the thread.

The gains over GLM 5.2 land in exactly those long-horizon jobs.

BenchmarkGLM 5.2GLM 5.3
SWE-Marathon19.442.5
Terminal-Bench 3.04.628.3
DeepSWE 1.146.266.9

Where it really shines is work that doesn't fit in one sitting: a migration that runs for hours, a repository too big to hold in your head, a security audit across a whole dependency tree. It holds a plan across thousands of steps instead of losing the thread halfway.

Pricing

Input / M
$0.50
Output / M
$2.00
Cached / M
$0.15

If you're already calling our API, the model name is the only thing that changes:

Model name
"zai-org/GLM-5.3"

Currently in beta

GLM 5.3 will stay in beta while we tune its performance. If you see unexpected responses, errors, or performance issues, post in our Discord. Your feedback goes straight to our engineers.

GLM 5.3 in Turbo

We serve it in Turbo mode on NVFP4. NVIDIA measured NVFP4 at or above FP8 on five of six benchmarks, so the speed costs you nothing in quality.

BenchmarkFP8NVFP4Δ
GPQA Diamond91.7992.68+0.89
SciCode58.6558.03−0.62
LiveCodeBench V652.9553.55+0.60
AA-LCR (long context)69.7569.63−0.12
IFBench65.3366.00+0.67
Terminal Bench 2.185.6786.80+1.13

Measured by NVIDIA on its GLM 5.3 NVFP4 checkpoint, served on SGLang.


#02IMPROVEDDEEPSEEK

DeepSeek V4.1 Flash now runs at avg. 220 tokens/s

We improved our inference and added capacity behind it. Same model, same price, twice the throughput: long generations finish in half the time, and agent loops that were waiting on tokens now spend that time working.


Get every update by email

Same news, sent to your inbox when it ships. Unsubscribe anytime.

All services are online

© 2026. Entrim. All Rights Reserved.

Privacy policy•Terms of service