The weird thing about this model is that it spits out Chinese characters just like the chinese models when chatting with it in German. Happened to me quite often during the past few days.
Opus 4.8 and Fable 5 did the same; they would output random Chinese characters and then say “I made a typo, let me fix it” and it’s just them removing that stay Chinese characters. Haven’t seen this since 5.1 family of models.
Does this suggest Mistral is distilling Deepseek et. al.? Can distillation keep passing through with high fidelity, or is it lossy like copying a VHS tape?
It doesn't, even remotely. It's a typical failure mode in all models, some do this more often. It's not related to samplers either. Not necessarily Chinese characters either, they likely just occupy a large portion of vocabularies. For example Gemini 2.0 used to glitch in Telugu and Ukrainian, Claude 2.x and 3.0 used to switch between random languages for a few words when writing in non-English, once in a blue moon.
No. I've had pretty much every model I've worked with (list too long to recite) spit out the odd Chinese, Korean, Thai or other Asian chunk of letters from time to time. ChatGPT in various incarnations has suddenly switched to Chinese in mid-output and remained that way until told to switch back. The only one I haven't really seen this from as far as I can recall is Gemini, but then again, I haven't used it as much because it was performing quite poor for what I was doing.
Ha, ha. Just like when Tilly Norwood switched to Chinese while interviewing with Piers Morgan [1].
On a serious note, I remember getting spooked when I saw a bunch of Chinese characters dumped out into some of my server logs (pre-LLM era). When I pasted the Chinese content from my log into google translate, it was just complete nonsense. Later I found a bug that was dumping raw binary data into my log file which interpreted as Chinese characters when viewed in a text editor. Perhaps similar bugs existed in the models.
I've seen this quite a bit on bad quants as well. I ran an early and very lossy quant of one of the Gemma3 models, and it loved to just stick Chinese and Korean characters into the text, sometimes without seeming to realize it had mixed languages.
Can you please share:
- what specific input you see this issue with?
- which interface you used (API, playground, specific sdk version or a wrapper app)?
- whether the chinese chars appear in the thinking section or the output?
Sorry it was not one time but it happened easily 5-10x during the past few days on different topics and prompts and I didn't store them. It felt like when pushing the model on some specifics in some niche topics in German (I have been prompting in German and the model responded in German). It was the acutal output and not the thinking traces and it was through the API. I jokingly asked it what the character meant and it didn't even make sense in the context so it seemed like random chinese characters.
There were a lot of people born in Russia with German ancestors and they had to leave Russia during WW2 or later left by themselves when the Soviet Union collapsed. Many chose to go back to Germany. This is also why there are probably a lot of Kazakhs in those statistics.
When the host dies with local storage, the VM is dead. With EBS it's not, since it simply can be booted up on another host. EBS is also replicated accross multiple hosts.
Thank you for your kind answer. I understand EBS as, in other words, virtualized storage, so then somewhere the logical block addresses have to be connected to actual physical storage devices, right? Then where and how is that mapping implemented? If you just give me the keywords, I’ll search for them myself.
Since I’m a freelance programmer, I mainly work on small-to-medium-sized projects, so I don’t know much about this kind of infrastructure programming.
Look up “SAN” or “storage area network.” EBS looks like a regular, local block device, another disk in /dev, but is actually served over the network. The implementation is transparent.
> Look up “SAN” or “storage area network.” EBS looks like a regular, local block device, another disk in /dev, but is actually served over the network.
When I think of "SAN" I think of Fibre Channel, which can be quite fast. EBS sounds more like iSCSI.
Anthropic trained their LLMs with copyrighted stuff but distillation is bad. Anthropic has closed models, China releases as open weight. DeepSeek even allows distilling their models, but China = bad. Understood.
Mistral has solid OCR, STT and TTS models and I would love to support them by switching with all of our business workloads to Mistral... but their LLM models are sadly not competitive at all. In our business benchmarks their Mistral Medium 3.5 with reasoning is worse than Gemma 4 31B and Glimmer 30B. It's a 128B dense model that's priced accordingly! Mistral Small 4 is way worse than Gemma 4 26B A4B. I applaude the effort that they release those models as open weight but Gemma 4 models are currently way easier to run with more tok/s and less hardware. Their API pricing is just insane for what you get. But I guess enterprise customers don't care about it, this is why they are probably not lowering it.
> Mistral Small 4 is way worse than Gemma 4 26B A4B
Depends for what purpose? I found large Mistrals are massively better than Gemma 4 at creative writing: have more natural tone, better consistency than 26B as it is MoE.
Mistral Small 4 is also a MoE model, with way more params, so I would expect it to perform better. Our benchmark involves around 10% communication and writing in German and English and Gemma 4 26B A4B beats it although creative writing is the only category in our benchmark which might be a bit subjective.
For me Opus 5 was the nail in the coffin. Fable without the restrictions was a great model, but became unusable with the security guardrails. Opus 5 became so bad and slow it's unbearable. And since Fable falls back to Opus all the time it was time to switch. Me and my friends are calling it Slowpus by now... Up until some weeks ago Anthropic was the king, but we all switched to Grok 4.6. It became slow as well but is not down all the time, is a solid model and paired with another reviewer model it's a great daily driver.
How long have you been testing this? Have you noted a large improvement? I tested Opus for this quite a while ago (maybe 4.5? Whatever was out about a year ago), and it performed quite poorly on my use case.
I have put together an internal benchmark on 1000s of business documents with weird tables, structure, etc. that I run on every relevant model release. Opus 4.8 performs very very well. But it is obviously overkill for the task (and expensive at doing so). I just wanted to respond to the OP.
I'm assuming that the reason I didn't have good success rate is because it was not scanned documents, but photographs, and lighting conditions weren't always ideal. I think scanned business documents are a happy-case scenario in a way. (obv, you seem to run it against some complex documents, so that's impressive)
reply