We're live on OpenRouter with DeepSeek v4.1-1017
Clear improvements on agentic coding across the board
Clear improvements on agentic coding across the board
DeepSeek V4.1 Flash 1017 is live on OpenRouter with DeepSeek as the first provider and many more providers on the way.
@deepseek_ai's update to V4.1 Flash is more token efficient and shows major gains in agentic benchmarks.
Use it now: openrouter.ai/deepseek/deeps…
0
3
37
We're live with deepseek-v4.1-flash-1017
🚀 DeepSeek-V4.1-Flash-1017 is here!
Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API.
🛠️ Optimized for @DeepSeekHarness, Tencent CodeBuddy, grade to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model.
1/4
0
0
11
Morph 🤝 SGLang
We're working with the @sgl_project and @lmsysorg team to push open-model disaggregated inference faster together.
First up is Kimi K3 Fast, served at up to 100 tokens per second through Morph's OpenAI and Anthropic-compatible APIs.
This is the beginning of a ound low-level optimization and high-performance open-model serving. More soon.
We're working with the @sgl_project and @lmsysorg team to push open-model disaggregated inference faster together.
First up is Kimi K3 Fast, served at up to 100 tokens per second through Morph's OpenAI and Anthropic-compatible APIs.
This is the beginning of a ound low-level optimization and high-performance open-model serving. More soon.
3
2
31
morph has signed the open weights letter
open weights matter because the gradient of capitalism corrupts absolutely in the absence of fair competition
thanks to @Microsoft for starting this and @JensenHuang for spreading it
open weights matter because the gradient of capitalism corrupts absolutely in the absence of fair competition
thanks to @Microsoft for starting this and @JensenHuang for spreading it
0
0
10
Reflexes are built to run on 100% of production amd run in under 90ms
Add more Reflexes to the same request with near-zero additional latency
How? We utilize shared backbone, multi head inference with a modern LLM
morphllm.com/blog/reflex-in…
Add more Reflexes to the same request with near-zero additional latency
How? We utilize shared backbone, multi head inference with a modern LLM
morphllm.com/blog/reflex-in…
1
0
3
2) Autoresearch for Kernels
H200 or B200 typically comes with NVLink
but weirder hardware with slow interconnect, poor uptime typically comes at a discount. we write kernels and systems to be reliable in these setups.
@tejasybhakta on July 1st: x.com/tejasybhakta/s…
H200 or B200 typically comes with NVLink
but weirder hardware with slow interconnect, poor uptime typically comes at a discount. we write kernels and systems to be reliable in these setups.
@tejasybhakta on July 1st: x.com/tejasybhakta/s…
excited to be speaking at @aiDotEngineer on July 1st about how we use autoresearch to discover kernels for weird hardware.
we're making models faster in environments that didnt make sense to buy until now
come listen
1
0
6
@peteyburn @albfresco @dhruvbhatia0 @mots_pod there are hooks for it but we're looking into it. There's no (public) interface for modifying context
1
0
4
Cheaper, Faster, and Better
#1 on SWE Bench Pro
We'll release traces of our SWE-Bench Pro Eval soon
#1 on SWE Bench Pro
We'll release traces of our SWE-Bench Pro Eval soon
Introducing WarpGrep v2. #1 on SWE-Bench Pro
What if code search was a separate, specialized model?
w/ WarpGrep v2: #1 on SWE-Bench Pro
Opus 4.6 → 57.5 (+2.1)
Codex 5.3 (cli) → 59.1 (+3.1)
MiniMax 2.5 → 57.6 (+3.7)
without WarpGrep:
Opus 4.6 (high) → 55.4
GPT-5.3 Codex .5 → 55.4
4
0
23
Findings after training for 15k+ b200 hours :
- RL on parallel tool formats lead to regressions on normal sequential tool use
- Some (but not all) tasks improve perf when split out into another model
- surprisingly hard to get models to not cheat evals with web search/git expensive
- RL on parallel tool formats lead to regressions on normal sequential tool use
- Some (but not all) tasks improve perf when split out into another model
- surprisingly hard to get models to not cheat evals with web search/git expensive
1
0
22
Splitting code search into a subagent with WarpGrep keeps the main models context clean.
instead of filling up context with junk while searching, our fast parallel subagent does it and returns ONLY the relevant content to the main agent.
This leads to gains for EVERY model.
The ance tank past 100k tokens - even 1M variants. Its critical to keep the main agents context clean for best perf
instead of filling up context with junk while searching, our fast parallel subagent does it and returns ONLY the relevant content to the main agent.
This leads to gains for EVERY model.
The ance tank past 100k tokens - even 1M variants. Its critical to keep the main agents context clean for best perf
3
1
26
Introducing WarpGrep v2. #1 on SWE-Bench Pro
What if code search was a separate, specialized model?
w/ WarpGrep v2: #1 on SWE-Bench Pro
Opus 4.6 → 57.5 (+2.1)
Codex 5.3 (cli) → 59.1 (+3.1)
MiniMax 2.5 → 57.6 (+3.7)
without WarpGrep:
Opus 4.6 (high) → 55.4
GPT-5.3 Codex .5 → 55.4
What if code search was a separate, specialized model?
w/ WarpGrep v2: #1 on SWE-Bench Pro
Opus 4.6 → 57.5 (+2.1)
Codex 5.3 (cli) → 59.1 (+3.1)
MiniMax 2.5 → 57.6 (+3.7)
without WarpGrep:
Opus 4.6 (high) → 55.4
GPT-5.3 Codex .5 → 55.4
44
27
331
WarpGrep v1 evals on real production code search:
WarpGrep: 0.73 f1 (6s)
SWE-Grep: 0.72 f1 (5s)
haiku: 0.72 f1 (73s)
haiku in the CC explore agent is not usable for live dev work because of how long it takes. swe-grep is fast on cerebras, but we're able to use new inference tectures on b200 to close the speed gap
WarpGrep: 0.73 f1 (6s)
SWE-Grep: 0.72 f1 (5s)
haiku: 0.72 f1 (73s)
haiku in the CC explore agent is not usable for live dev work because of how long it takes. swe-grep is fast on cerebras, but we're able to use new inference tectures on b200 to close the speed gap
3
0
28
watch all of your PRs
Introducing Glance, a fast browser agent trained with RL to test code changes.
We're releasing this with BrowserBot, which puts a video of it testing your app right in your GitHub PRs
Frontier models are trained to be helpful. Our model is trained to be bad at using software.
3
6
99
We're also thrilled to be shipping with first class support for @vercel △ !
Introducing Glance, a fast browser agent trained with RL to test code changes.
We're releasing this with BrowserBot, which puts a video of it testing your app right in your GitHub PRs
Frontier models are trained to be helpful. Our model is trained to be bad at using software.
2
1
18
How to install this on all of your Github PRs in under 30s :
Introducing Glance, a fast browser agent trained with RL to test code changes.
We're releasing this with BrowserBot, which puts a video of it testing your app right in your GitHub PRs
Frontier models are trained to be helpful. Our model is trained to be bad at using software.
0
2
18
Why train this? We turn code testing into a search problem – searching for components that changed via browser tools.
The hard part is grounding. A diff in Button.tsx doesn’t tell you where that button lives or how to reach it.
We walk React’s Fiber tree to map changed files → boxes.
Given any diff, we know exactly which pixels need testing. Tools like Bippy and react-grab from @aidenybai were useful starting points.
The hard part is grounding. A diff in Button.tsx doesn’t tell you where that button lives or how to reach it.
We walk React’s Fiber tree to map changed files → boxes.
Given any diff, we know exactly which pixels need testing. Tools like Bippy and react-grab from @aidenybai were useful starting points.
2
0
20
Introducing Glance, a fast browser agent trained with RL to test code changes.
We're releasing this with BrowserBot, which puts a video of it testing your app right in your GitHub PRs
Frontier models are trained to be helpful. Our model is trained to be bad at using software.
We're releasing this with BrowserBot, which puts a video of it testing your app right in your GitHub PRs
Frontier models are trained to be helpful. Our model is trained to be bad at using software.
14
7
114
Add morphllm [dot] com/ infront for the github in the url and ask away
We have open-sourced our new 𝕏 algorithm, powered by the same transformer architecture as xAI's Grok model.
Check it out here: github.com/xai-org/x-algo…
3
0
22
We found that giving models access to warp grep boosts performance between 5-12% with frontier models, while improving speeds and reducing token usage up to 40%
WarpGrep is our attempt to fix coding agents.@swyx
hit the nail on the head with the semi async valley of death. eeps developers in flow by making search 10x faster, and reducing context rot by 50+%
WarpGrep is our attempt to fix coding agents.@swyx
hit the nail on the head with the semi async valley of death. eeps developers in flow by making search 10x faster, and reducing context rot by 50+%
2
1
46
Context pollution does more than just waste tokens, it actively kills model performance.
In production repos it poisons the model’s thinking into editing files you didn’t want to touch, while costing exponentially more.
The WarpGrep approach: have a inference optimized model through repos for you.
WarpGrep is smart enough to understand codebases, and RL’ed specifically to ingest a ton of data and act with high recall on extra long context.
In production repos it poisons the model’s thinking into editing files you didn’t want to touch, while costing exponentially more.
The WarpGrep approach: have a inference optimized model through repos for you.
WarpGrep is smart enough to understand codebases, and RL’ed specifically to ingest a ton of data and act with high recall on extra long context.
1
1
54
Coding Agents spend 60% of their time searching instead of coding.
We built WarpGrep so people can have sub 10 second agents, without compromising on accuracy.
Dev flow-state is important. P(bounce) increases 10% with each second a dev has to wait according to @swyx and cognition research
We built WarpGrep so people can have sub 10 second agents, without compromising on accuracy.
Dev flow-state is important. P(bounce) increases 10% with each second a dev has to wait according to @swyx and cognition research
1
6
101
Introducing WarpGrep, a fast context subagent that improves coding agent performance.
WarpGrep speeds up coding tasks 40% and reduces context rot by 70% on long horizon tasks by treating context retrieval as its own RL trained system.
Inspired by Cognition’s SWE-Grep - we’re e Code, Codex, OpenCode or any coding agent via MCP (or through our SDK)
WarpGrep speeds up coding tasks 40% and reduces context rot by 70% on long horizon tasks by treating context retrieval as its own RL trained system.
Inspired by Cognition’s SWE-Grep - we’re e Code, Codex, OpenCode or any coding agent via MCP (or through our SDK)
67
88
1,075
Search is the bottleneck in every coding agent.
We found that adding semantic search makes coding agents 31% more accurate on large codebases
Morph now supports Semantic Search built directly into your Git repo.
Try it now→ docs.morphllm.com/sdk/components…
We found that adding semantic search makes coding agents 31% more accurate on large codebases
Morph now supports Semantic Search built directly into your Git repo.
Try it now→ docs.morphllm.com/sdk/components…
2
1
12
Frontier coding models are expensive. We’re releasing a new Model Router that makes them cheaper without compromising
Try it through our brand new SDK:
docs.morphllm.com/sdk/components…
Try it through our brand new SDK:
docs.morphllm.com/sdk/components…
1
2
17
Cognition nailed it with the “Semi-Async Valley of Death.”
If your agent is too slow to keep you in flow but too inaccurate to run autonomously it’s dead.
We built Morph Fast Apply to escape that valley. It applies code edits at 10,500 tokens/sec and hits near-perfect acy is speed.
Retries kill latency. Flow breaks with every second lost.
Small, specialized models that delegate fast + precise work are the future of agentic AI.
Fast Apply is our first step toward that sub-agent world. Stay tuned for more!
If your agent is too slow to keep you in flow but too inaccurate to run autonomously it’s dead.
We built Morph Fast Apply to escape that valley. It applies code edits at 10,500 tokens/sec and hits near-perfect acy is speed.
Retries kill latency. Flow breaks with every second lost.
Small, specialized models that delegate fast + precise work are the future of agentic AI.
Fast Apply is our first step toward that sub-agent world. Stay tuned for more!
2
1
12
The Agents era just got real.
@MetorialAi Agent Jam Hackathon killed it this weekend at the @ycombinator office! Was so great to see how all the teams used Morph in their products.
we saw 6 teams build agents that actually do work: call people, deploy apps, manage ops, train
🏆️Finalists below ↓
@MetorialAi Agent Jam Hackathon killed it this weekend at the @ycombinator office! Was so great to see how all the teams used Morph in their products.
we saw 6 teams build agents that actually do work: call people, deploy apps, manage ops, train
🏆️Finalists below ↓
7
8
25
They're going to sunset Supermaven next month
This is my 9/11
@cursor_ai mark my words, I will NEVER forgive you for this
0
0
8
