Last two weeks were action packed. Here are the changes since then:

• (Boosted+) DeepSeek V4.1 Flash boosted from $60 usage to $80 usage!
• (Boosted++) DeepSeek V4.1 Flash has been updated to latest 1017 with an incredible inference boost
• (Removed) DeepSeek V4 Flash and Vision, makes sense as both are obsolete now
• (New) Kimi K3.1 Pro flagship Kimi model
• (New) MiniMax M3.1 Flash looks like a respectable leap form M3 at cheaper token prices
• (New) GPT-6 Luna cheaper than 5.6 Luna but lower Go credits, and roughly performs at the same level
• (New) Gemini 4 Argon contributor want to feed google more of your data? here's your chance! 😅

With the latest DeepSeek V4.1 Flash upgrade (1017), according to AA scores (+5 points), flash is now at the same level as GLM-5.3, Qwen3.8-Max, and Grok 4.7! how insane is that.

MiniMax M3.1 Flash offers less credit than M3 but it's cheaper so they end up roughly equivalent quota-wise.

Kimi K3.1 Pro unfortunately shows signs of bench maxxing gone too far in addition to cybersecurity nerf, and to top it off there's a self-inflicted PR wound on Kimi's side so not the smoothest release of all time. All that being said, it still might be the highest performing open weights model to date.

Note that I don't report the named secret models, temporary previews, and other shenanigans. Also in some cases I remove older obsolete models from the chart, for example MiMo-v2.5 or MiniMax-M3.

More and more this feels like I'm reporting on some organized sports League, ups and downs, wins and loses, and the occasional scandal every now and then 🏆

Been almost a month since the last update, now that the dust of August and September's LLM insanity is cleared out, new models are settling in nicely.

A number of updates over last week or so:

• (Boosted) DeepSeek V4.1 Flash switched from timed 4x usage to permanent $60 usage
• (New) MiMo-V2.6 Pro and MiMo-v2.6 Flash were both added with identical limits to V2.5
• (New) Claude Haiku 5.5 was just added but it's more expensive compared to its direct competitor GPT-6 Luna (both $15 credit)

MiMo-V2.6 in benchmarks is a massive leap from MiMo-V2.5 for both Pro and Flash, and it's definitely better no denying it, but I noticed a significant disconnect between real world performance (coding) vs what the benchmarks portray. Sadly MiMo-V2.6 appears to be blatantly bench-maxed (too much) that expectations don't match reality immediately.

Currently for me I like DeepSeek V4.1 Flash for most soft skills like writing automation and researching ideas and it also runs my hermes agent's and in my opinion has the best personality and is blazingly fast (often above 200 t/s!).

GLM-5.3-Flash is my go to for coding, although sometimes Qwen3.8 Flash actually performs better or on par even though its benchmarks put it slightly below, I still perceive GLM-5.3 LLMs as ace for coding.

After the insanity of August where one specific week saw the release of 10 new exciting LLMs, September has been calm in comparison (so far).

Updates in OpenCode Go:

• (New) DeepSeek V4.1 Flash: next big DeepSeek moment is here. Soft replaces DeepSeek V4 Pro. Fast and affordable with vision. $15 usage only expected to double at least
• (New) Muse Spark 1.3 Contributor: the next iteration of Muse Spark, nearly free but Zuck gets all your data and meta geo-bans most of the world
• (Leveled Up) the excellent GLM-5.3-Flash was raised from $15 to $60 credits permanently, 4x the usage quota. This alone restores OpenCode Go's great bargain status.

One major update that OpenCode Go didn't get yet as far as I know is a stealth Qwen3.8-Max significant upgrade from a single tweet and no real announcement, the new version is named Qwen3.8-Max-0902 and pushes Max's capabilities to past Kimi K3 and puts it where Fable/Opus5 are. Qwen and OpenRouter already provide the new version. The benchmarks for it aren't even on Artificial Analysis yet.

This week's special Go multiplier is the new DeepSeek V4.1 Flash at 4x. Incredible that a model at the level of Opus5 is essentially free for a while.

OpenCode Go began sending automated emails notifying affected users a few days ago. The email they sent might have ended up in the spam folder if you use gmail.

Starting 6th of September, requests to OpenCode Go will fail if they didn't contain the session header.

The required header is pretty straight forward: {'x-opencode-session': '<stable-id-per-conversation>'}

They require this header to better manage tokens caching on third party harnesses/services.

One specific affected harness is Hermes, it needs to be upgraded past v0.21.0 to include that header.

DeepSeek Harness is another affected one, but it seems opencode updated their logic on their end to also support the deepseek-specific session ids based on this post while still requiring the header so the fix might be temporary.

I created this chart to help me quickly pick the best value model at the moment. This chart is updated to current Go plan status (2x GLM-5.3-Flash).

The blue bar represents usage limits of that model per week, the red line is the Artificial Analysis Intelligence index of the model sorted from highest to lowest.

At this moment, DeepSeek V4 Flash and Qwen3.8 Flash are the best models to use for the majority of cases. GLM-5.3 -Flash is great for more demanding work.

The GOATs have arrived.

Both very affordable and better than DeepSeek V4 Pro and GLM-5.2 (if benchmarks mean anything at this point).

• (New) GLM-5.3-Flash: competitive with GPT-5.6 Sol (high) with bonus vision! even at only $15 of usage, it's still affordable
• (New) Qwen3.8 Flash: competitive with GPT-5.6 Terra (max) and also with vision. $30 usage from the get go (and fast)
• (New) Hy4 Preview: Tencent's first frontier model, needs more work but looks very capable already. $30 usage

Remember how happy we were with DeepSeek V4 Flash 0731? this inference and capability acceleration of local-friendly open weights models is a little scary..

I'm personally most impressed by Qwen3.8-Flash-Next, this has a new architecture meant to power Qwen4 that allows it to partially utilize SSD storage to offload about 30% of its parameters (n-gram weights). Even though this model is supposed to be for preview/research, it's already capable enough to beat GPT-5.6 Terra (xhigh). I predict the first Qwen4 Max will at least match Opus 5!

Imagine a new 27B or 35B using this same architecture..! only one way to go from here 🏹

Yesterday I installed the deepseek harness after hearing a lot of noise about it from different channels. It's minimal but with the idea that everything is a plugin and agents can create plugins dynamically so it basically has infinite flexibility.

Today I decided to get it to build a human/agent collaboration pipeline. For each workspace (project), an AGENTS.kanban is created that has a kanban board formatted as json to match the vscode extension portable-kanban. The board is fully viewable and functional in the harness web ui (as well as vscode) with an identical set of features to portable-kanban. Then I can use the command: /kanban next to instruct the agent to pickup the next todo task from the Doing or To Do boards and implement it.

To implement all this I used Ornith-1.5-35B running on my AI PC (with a modest 32+8 memory), and when there were complex issues or the harness was throwing errors, I enlisted the help of GLM-5.3-Flash through the lite qwen token plan ($6!). I ran into two issues that caused previous dsh sessions to fail to parse both were debugged and fixed by GLM.

Now I'm at end of day, I have a fully functional kanban collab plugin which I used to develop the plugin itself. The agent also started contributing new tasks into the backlog (ones marked with "Confirm" label).

I'm not sure where I'll go next with it, but it's already usable! which is astonishing.

My experience with deepseek harness so far:

The Good

• Coming from opencode, dsh is superior in its clarity and runtime loop
• It's noticeably lighter to run than opencode on my system despite my running it through the web interface (dsh web) rather than TUI
• The everything is a plugin philosophy is a great approach imo
• Using workspaces as a first class citizen, workspaces map perfectly to projects and all are active and accessible always. So I can launch tasks on 2+ projects concurrently from the same interface

The Bad

• Opinionated towards DeepSeek API. I still don't know how to turn off their default web search tool from settings. It requires a DeepSeek API key and fails when models call it if a key is not provided. Even when I setup my own MCP for web search, models default to attempting to use the built-in tool and fail every time (then tries the mcp tool).
• Not clear which parts of the harness can be changed on the fly and update immediately, and which parts require a full harness restart to apply.
• Has the same issue as opencode: uses a hardcoded list of models per provider rather than fetching the model list from v1/models endpoint; leading some providers to show a stale list of models (such as qwen's token plan).
• Functionality breaking bugs but that's expected

---

If you've tried it, what's been your experience with deepseek harness?

Hy4 preview is a new-generation Mixture-of-Experts (MoE) flagship model developed by the Tencent Hy Team. The model comprises 770B total parameters, of which 49B are activated per token. Model supports 1M tokens context length.

• Hy4-Preview on HuggingFace
• Hy4-Preview on ModelScope

Native FP8 weights are published as well.

Benchmarks:

Known issues with the preview model: spends longer than necessary reasoning through complex tasks, and a tendency to over-verify its own work.

Open-weights frontier models will continue to drop until RAM prices improve 😅

• (New) Grok 4.6: latest frontier level Grok at only $15 usage
• (New) LongCat-2.0 this is a last gen weak open-weights model
• (Removed) Grok 4.5

LongCat-2.0 is a weak 1.6T-A48B (same size as DeepSeek V4 Pro btw) so it's bewildering why they can serve this one at such generous quota but kneecap DeepSeek V4 Pro.

LLM token economics make no sense, they might as well be stocks at this point

• (New) DeepSeek V4 Flash Vision Exp: new experimental DeepSeek V4 Flash build, adds vision and a small bump in inference as well
• (New) Ox Alpha Free: entirely free and 100% unlimited, only for 1 week though. No one knows what it actually is but it's very capable and multi-modal (equivalent quality to GLM-5.3 and GPT-5.6 Sol)

Model                         5HOURS  1WEEK   1MONTH    CREDITS
-----------------------------------------------------------------
DeepSeek V4 Flash Vision Exp  3800    9450    18900     $15
0x Alpha Free (secret test)   -       -       -         unlimited

0x Alpha is better than DeepSeek V4 Flash 0731 and OpenCode has significantly more capacity for it (100T tokens/day). Sources indicate this model is going to be a surprise when revealed. My guess? a marketing stunt for a new Google or OpenAI model. Reason why I say this is that the model performance is 100% frontier-level and capacity provided is impressive. Whoever is doing this already has a significant infrastructure in-place that's not used much 😅 which excludes most Chinese providers.

New changes in opencode Go plan:

• double DeepSeek V4 Flash usage quota raising it to a $30 credit (from $15)
• add Muse Spark 1.2 Contributor, this model is on par with GPT-5.6 Terra (based on artificial analysis score) and has a very generous quota. However it sends all your data to Meta for collection 🫤

Model                       5HOURS  1WEEK   1MONTH
----------------------------------------------------
Muse Spark 1.2 Contributor  45,300  113,300 226,600
DeepSeek V4 Flash           7,600   18,900  37,800

OpenCode Go plan which STILL advertises that it gives users $60/month of credit for each $10/month sub making it one of the best LLM plans out there, is no longer actually doing that. They are now giving users a single pool of $15 of credits for each $10/month sub for all their new models: GLM-5.3, Kimi K3, DeepSeek V4 Pro 0813, and DeepSeek V4 Flash 0731.

This is in spite of new DSV4 and GLM models actually remaining exactly the same as the previous versions but with post-training (same hardware and memory requirements), there's no practical reason for these models to increase prices from both OpenCode and DeepSeek other than sacrificing their goodwill and userbases to increase profit.

This especially applies to DeepSeek V4 Flash as its one of the least demanding on hardware. Some older models still offered by the opencode Go plan at $60 credits require significantly more hardware! (MiMo-V2.5 is a larger model for example)

And the worst part is that opencode still (til this moment) write in their Go FAQ that they offer $60/month credits so this is less than honest on their part:

This is an MoE model with 1.6T-A49B

The weights were up briefly then taken down due to some issues in the repo files apparently, now they're back up:

• ModelScope
• HuggingFace

GGUFs are out as well:

• Unsloth's DeepSeek V4 Pro 0813

DeepSeek published benchmarks for reference:

GLM-5.3 is a post-trained GLM-5.2, still 753B parameters.

Unfortunately, usage cost of GLM-5.3 in Go is currently 4x GLM-5.2, but I expect they'll adjust that once the open weights drop in ~2 weeks:

• 220 requests per 5 hours (880 for GLM-5.2)
• 540 requests per week (2150 for GLM-5.2)

For me, I've set up two additional agents besides Plan and Build:

• Learn: rather than try to implement everything, responds with code snippets and guidelines to implement the feature I want. This speeds up my coding without taking away my agency over the code or building experience. It has become my favorite go-to mode.
• Reviewer: what it says, reviews the code, categorizes its conclusions into: Critical, High, Medium, Low, Polish. Also doesn't modify the code, just provides brief guidelines to resolve the issue.

Pipeline

Generally I start by describing the feature to the Learn agent, it explores the code and creates instructions/guidelines for me to implement it that account for the existing code base expectations and behaviors. I just dive in and implement it.

Then I use the Plan and Build modes to create test suites for the feature as well as implement any trivial code needed.

Once all is clear and ready, I ping the Reviewer to do its thing and usually there are a few false positives, but it does catch real issues. I then implement any necessary fixes and iterate until code is ready.

Model assignments

• Learn: should assign a pro level model with a good linguistic personality. so far I like using: MiMo-V2.5-Pro, Terra-high, DeepSeek V4 Pro
• Plan: should be a frontier-level model. My gotos are: Sol-high and GLM-5.2
• Build: any good coding model. My current gotos are: DeepSeek V4 Flash 0731, Terra-high, Luna-high
• Reviewer: some frontier-level models (like Sol) aren't good for this because they flag way too many "what-if" issues. I prefer: GLM-5.2, Terra-xhigh, DeepSeek V4 Pro

Under the name: DeepSeek V4 Pro (New)

If you can't see it, you probably need to turn on the chinese llm providers toggle in Go's control panel.

DeepSeek V4 Pro 0813 is already available in OpenRouter costing the same as the preview: $0.435 /M tokens input, $0.87 /M tokens output

Published benchmarks puts it way ahead of GLM-5.2, and close to Kimi K3, definitely at frontier level.

As a reminder, Fable 5 costs: $10 /M tokens input, $50 /M tokens output. Making DeepSeek V4 Pro is 20-50x cheaper!

opencode Go is expected to provide access to the new model within 24 hours like they did for previous model releases.

I could not find a community for opencode in lemmy so decided to create one.

Reddit enshitified years ago and deserves to be left behind for something stable and open.

So welcome here stranger, pick up an LLM and sit around the agentic fire.

Glad they set a time rather than leaving it open

In the release page: "The other model(s) in Qwen3.8-series would be released later", we know that Qwen3.8-27B is dropping next, but that statement implies there might be other models beyond the 27B!

• Qwen3.8-Max on ModelScope
• Qwen3.8-Max on HuggingFace