Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

MAI-Code-1-Flash

microsoft.ai

Post reply on HN
Print view View original

191–200 of 297 posts

Re: MAI-Code-1-Flash

#191

Post by verdverm »

chillfox wrote: post #184
verdverm wrote: post #94

Earlier quoted context omitted.

I've been having really good results with DeepSeek-v4-flash, qwen-3.6-moe, and the older gimini-3-flash-preview. (recent geminis suck hard) Small models are more than enough for the majority of tasks these days. Plan and review with the bigger ones, let the little ones explore and implement. OpenCode Go is $10/month for the open weight models with nice quotas: https://opencode.ai/go

You don’t have to limit yourself to the tiny models with the OpenCode Go plan, you can get a lot of usage from the bigger models if you keep the cache hot. I am about 85% through my quota with 9 days left before refresh and have just used over 1B tokens, mostly DeepSeek V4 Pro, but also a little mimo 2.5 pro and kimi k2.6

For sure, I've been flipping between flash/pro (or the equivalent for other families), been trying to stick to one family per project as a way to test them out independently over longer periods and more realistic/diverse tasks. I've definitely spent more quota on pro and pushed more tokens through flash.

Re: MAI-Code-1-Flash

#192

Post by lemonish97 »

jnwatson wrote: post #189

I had to remind myself what Haiku is even for. Anthropic hasn't spent a lot of recent marketing on it. When I need a light model, I reach for Sonnet. It is nearly free on the max plans, and quite fast. I don't see a place for Haiku in regular coding. Haiku I guess is when you need summarization/categorization at scale. Microsoft setting Haiku as the benchmark is a low bar.

> "It is nearly free on the max plans"

is a funny oxymoron

Re: MAI-Code-1-Flash

#195

Post by veber-alex »

tedggh wrote: post #126
hparadiz wrote: post #106

Earlier quoted context omitted.

The $20/month ChatGPT plan that comes with codex is good value. Even just have premium ChatGPT is nice. I get rate limited regularly but it still lets me do most things.

The $100/month is excellent value. I don’t understand how’s that not the default option for all professional developers. Unless people don’t produce any value writing code, like playing around and experimenting with vibe coding, I understand. But if software development is your actual income, and assuming you live in a wealthy country, $100/month is nothing for a tool like Codex.

Every developer who writes code for a living should get an AI subscription from work and not have to pay for it himself.

Re: MAI-Code-1-Flash

#196

Post by treesknees »

papascrubs wrote: post #177
GaryBluto wrote: post #174

Earlier quoted context omitted.

Using OpenRouter, going to migrate to DeepSeek's official API soon. I'm not using it for anything commercial or for private data so I have no privacy qualms.

Makes sense. Privacy is my only real hang up with DeepSeek. Both of the big SOTA providers have become extremely filtered. Things that I could do one version ago are now getting refusals. Anthropic is almost unusable. ChatGPT is slightly better. Even with a "cyber exception" in place and a vetted account. They are going to force me to take my business elsewhere.

GitHub Copilot refuses to do any security testing or proof-of-concepts for exploits. While I understand why, we pay for Enterprise and I’m working on our proprietary code base. It’s incredibly annoying.

I’ve actually had luck taking the analysis from GHCP and pasting it into our M365 Copilot and getting a useful poc to stick into my bug reports.

Re: MAI-Code-1-Flash

#198

Post by easygenes »

camelmel wrote: post #43

Huh, according to that model card this is a 137B total parameter model. Performance doesn't seem that good: - MAI-Code-1-Flash (137B-A5B) = 51% on SWE-bench pro - Qwen3.6-35B-A3B = 49.5% on SWE-bench pro ( https://huggingface.co/Qwen/Qwen3.6-35B-A3B ) They benchmark against Claude Haiku but Haiku is not good, it's worse than tiny open models you can run locally or via API at 10% the cost.

While I agree directionally, I'll caveat that "cost per token" != "cost per task". In the case of Qwen3.6 it tends to think 1.6x more than Haiku, so the cost of Haiku on the same tasks tends to only be about double. More detail from comparing their Artificial Analysis metrics:

  Qwen3.6-35B-A3B   vs   Claude Haiku 4.5
    reasoning mode · AA Intelligence Index v4.0
  
  46.0 ┤   ↖ better — cheaper · smarter · faster
       │
       │
  44.0 ┤     ╭─────╮
       │     │  ●  │ Qwen3.6-35B-A3B
       │     ╰─────╯
  42.0 ┤
       │
       │
  40.0 ┤
       │
       │
  38.0 ┤                                       ╭───╮
       │                      Claude Haiku 4.5 │ ○ │
       │                                       ╰───╯
  36.0 ┤
       └┬─────────┬─────────┬─────────┬─────────┬────────┬
        $200    $300      $400      $500      $600    $700
  
    x → cost to run the index (USD)        lower is better
    y → AA intelligence index              higher is better
  
    bubble area = output speed (tokens / sec)
          ╭─────╮                  ╭───╮
          │  ●  │ Qwen ~196 t/s    │ ○ │ Haiku ~93 t/s
          ╰─────╯                  ╰───╯
  
    ┌─────────────────────┬──────────┬──────────┬───────────┐
    │ model               │ AA index │ run cost │ out speed │
    ├─────────────────────┼──────────┼──────────┼───────────┤
    │ Qwen3.6-35B-A3B    ●│   43.5   │   $280   │  196 t/s  │
    │ Claude Haiku 4.5   ○│   37.1   │   $620   │   93 t/s  │
    └─────────────────────┴──────────┴──────────┴───────────┘


    COST PER TOKEN   ≠   COST PER TASK  
    output tokens per index run:
       Haiku 4.5    87.3M   (79.3M reasoning + 8.0M answer)
       Qwen3.6     143.2M   (131.7M reasoning + 11.5M answer)
       → Qwen emits 1.64× more output
  
    ── output speed (tokens / sec) ──────────  raw rate · higher = faster
       Qwen3.6     100%   ~196 t/s
       Haiku 4.5   ~47%   ~93 t/s
                                                  → Qwen ~2.1× faster per token
  
          ╎   1.64× more tokens

Re: MAI-Code-1-Flash

#199

Post by roywiggins »

eli wrote: post #175
bel8 wrote: post #82

It's a start and I welcome competition but I don't think I ever used small cloud models like Haiku 4.5. They are cute but for serious coding they tend to waste your expensive time. And this certainly wont bring me back to GitHub Copilot which I cancelled yesterday. GitHub Copilot had competitive pricing until yesterday when they changed from per-request to one of the most expensive per-token quotas. Seriously, take a…

Makes sense as part of a larger coding workflow, especially if it’s fast. Using a trillion parameter model to figure out how to call a targeted edit tool or generate a commit message is a waste. Also narrow tasks like “make the background darker” or “rename this function and update callers”

> “rename this function and update callers”

I'm old enough to remember when IDEs could do this without needing a couple gigabytes of matrices to do it

(LLMs are great for anything even slightly more complicated ofc)

Re: MAI-Code-1-Flash

#200

Post by roywiggins »

swores wrote: post #140
stefan_ wrote: post #116

Earlier quoted context omitted.

Anthropic & co charge API users much more, not least to demolish the middlemen low-effort plays like Cursor and Copilot. To not own the model is not viable in 2026.

Sorry, what do you mean by "To not own the model is not viable in 2026." I assume I'm misunderstanding you (likely my fault), because the way I read that is that you're saying nobody should currently be using models owned & hosted by companies like OpenAI and Antheopic, while clearly a huge number of people are using those in 2026 despite not owning them.

It's that companies like copilot/cursor are in real trouble if they are in the business of reselling expensive Anthropic tokens
Post reply on HN